Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Evaluating deep learning EEG-based mental stress classification in adolescents with autism for breathing entrainment BCI

Evaluating deep learning EEG-based mental stress classification in adolescents with autism for... Mental stress is a major individual and societal burden and one of the main contributing factors that lead to patholo‑ gies such as depression, anxiety disorders, heart attacks, and strokes. Given that anxiety disorders are one of the most common comorbidities in youth with autism spectrum disorder (ASD), this population is particularly vulnerable to mental stress, severely limiting overall quality of life. To prevent this, early stress quantification with machine learning (ML) and effective anxiety mitigation with non‑pharmacological interventions are essential. This study aims to investi‑ gate the feasibility of exploiting electroencephalography (EEG) signals for stress assessment by comparing several ML classifiers, namely support vector machine (SVM) and deep learning methods. We trained a total of eleven subject ‑ dependent models‑four with conventional brain‑ computer interface (BCI) methods and seven with deep learning approaches‑ on the EEG of neurotypical (n=5) and ASD (n=8) participants performing alternating blocks of mental arithmetic stress induction, guided and unguided breathing. Our results show that a multiclass two‑layer LSTM RNN deep learning classifier is capable of identifying mental stress from ongoing EEG with an overall accuracy of 93.27%. Our study is the first to successfully apply an LSTM RNN classifier to identify stress states from EEG in both ASD and neurotypical adolescents, and offers promise for an EEG‑based BCI for the real‑time assessment and mitigation of mental stress through a closed‑loop adaptation of respiration entrainment. Keywords: Mental stress, Autism, EEG, Deep learning, Breathing entrainment 1 Introduction of stress even in the absence of these stressors. The Individuals with autism spectrum disorder (ASD) often comorbidity of anxiety disorders and ASD in children demonstrate deficits in social communication skills and and adolescents has been studied extensively with 40% restricted or stereotyped behaviors and interests [1]. This to 85% of individuals with ASD aged 6 to 18 having at causes those with ASD to experience states of cogni- least one form of anxiety [3–5]. Unfortunately, individu- tive and emotional overload, leading to increased stress als with ASD are uniquely vulnerable to the deleterious and ultimately anxiety symptoms [2]. Although there effects of stress because of their hyper- or hyporeactiv - is significant overlap between stress and anxiety, stress ity to sensory inputs, as well as difficulties with accurate is best understood as the physiological and psychologi- stress detection and coping with stressful situations [6]. cal response towards stressors; anxiety is the persistence Given the frequency in which anxiety co-occurs in ASD, in conjunction with the hurdles in education, long-term functional impairments, reduction in quality of life, and *Correspondence: adrien.martel@icm‑institute.org increased caregiver burden [7–13], a more comprehen- Cognitive Neuroscience and Information Technology Research Program, sive understanding of comorbidities in ASD as well as Open University of Catalonia (UOC), Barcelona, Spain personalized intervention methods to relieve clinical Full list of author information is available at the end of the article © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Sundaresan et al. Brain Inf. (2021) 8:13 Page 2 of 12 symptoms of the disease and improve emotional and and passive brain-computer interfaces (pBCIs) aim at physical well-being for individuals with ASD is needed. intelligent forms of adaptation in response to cognitive Incidentally, anxiety and the design of appropriate inter- state assessments [30, 31]. The field of EEG-based BCIs vention methods have been identified by the autism com - has blossomed in recent years, largely on account of munity and clinicians as a key priority with researchers EEG’s high temporal resolution, non-invasiveness, rela- emphasizing the need for more precise measures of anxi- tively low cost, and novel advances in the effectiveness ety [14]. Moreover, the lack of objective and continuous and usability of acquisition systems [32, 33]. While BCIs measurements of stress is particularly detrimental for have historically been employed in the context of assis- a population already affected by an inability to express tive technologies for severely impaired individuals [34, inner experiences and calls for novel methods to identify 35], pBCIs have mainly been aimed at developing adap- individualized stress markers in real-time [15]. Among tive automation for real-world applications [36–38]. The the triggers identified, such as challenging sensory expe - central challenge of EEG-based pBCIs is to account for riences or social demands, anxiety related to academic the high inter- and intra-individual variability of neuro- expectations is thought to have the greatest impact on physiological signals exhibited under particular cognitive school performance for ASD children and adolescents states [39]. However, by averaging over a large enough [16, 17]. number of samples it is possible to distill sufficiently spe - Concurrently, a growing number of studies have dem- cific brain activity patterns and train a machine learning onstrated the efficacy of stand-alone meditation, relaxa - classifier to learn to discern these patterns in real-time tion and breathing practices for improving well-being, [40, 41]. This approach has already been successfully mental health and managing stress [18–20]. Although applied to monitor several cognitive states such as work- the underlying mechanisms are not yet fully understood, load [42–44], vigilance [45–47], and fatigue [48, 49]. breathing practices such as ‘anatomically optimized res- Neurofeedback involves monitoring a user’s mental piration’, i.e. controlled, slow diaphragmatic breathing state with EEG and providing feedback through a vari- through the nose in the range of 6-10 breaths per min- ety of modalities (visual, audio, tactile, etc.) to modulate ute (brpm) or resonance frequency breathing, 4.5 to 6.5 particular biomarkers [50]. In conjunction with breath- brpm for adults and 6.5 to 9.5 brpm for children, have control, neurofeedback training has been shown to be been found to procure significant physiological benefits a promising mitigatory tool for anxiety. For example, [21, 22], reduce physiological and psychological stress White et  al. [51] demonstrated that breathing-based [23] and even improve sustained attention performance visual neurofeedback reduces symptoms in patients with [24]. Prior studies have shown that breath-control can anxiety and depression, while acoustically mediated deep address physiological correlates of anxiety, including breathing neurofeedback was shown by Crivelli et al. [52] heart rate variability [22, 25], a well-validated quantita- to diminish long-term stress and anxiety levels in young tive stress indicator [26]. Notably, breath-control has adults. The first step towards an EEG-based BCI able to been found to significantly decrease test anxiety in stu - monitor anxiety levels, identify an individual’s optimal dents in an educational setting [27]. Moreover, a recent breathing patterns, and adapt breathing entrainment review by Zaccaro et  al. [28] found that controlled slow parameters in real-time, is to determine whether mental breathing (<10 brpm) had a significant impact on auto - stress can be classified on the basis of ongoing EEG data. nomic nervous system activity, especially in the theta Classification algorithms are key elements of any EEG- (3–7 Hz), alpha (8–14 Hz) and beta (15–30Hz) bands based BCI’s ability to recognize users’ EEG patterns and of the electroencephalogram (EEG), linked to improved associated cognitive states. Among the large diversity of cognitive performance during attentional and executive existing architectures and types of classifiers (for reviews functions [29]. Although these findings, taken together, see [53] and [54]), deep learning methods have recently speak to the promise of using controlled slow breathing emerged as methods of analysis that can consider neu- as a simple, low-cost and non-pharmacologic interven- rophysiological data in its entirety, including the time tion [23] to mitigate anxiety, optimized efficacy hinges on domain [54]. Convolutional neural networks are the assessing an individual’s current level of stress and ideal most widely used deep learning algorithms in EEG analy- respiration parameters in real-time. sis [55], and have been shown to be effective in emotion Although cognitive or affective states such as stress are detection [56, 57] and anxiety classification [58] in par - not directly observable externally nor reliably measurable ticular. Further, deep learning with convolutional neural through behavioral measures or subjective reports, devel- networks (CNNs) have recently been shown to outper- opments in EEG-based brain-computer interfaces (BCIs) form the widely used filter bank common spatial pattern have increasingly permitted the continuous and real-time (FBCSP) algorithm [59] by extracting increasingly more monitoring of mental states. Neuroadaptive technologies complex features of the data [60]. Accordingly, we aimed Sundar esan et al. Brain Inf. (2021) 8:13 Page 3 of 12 at comparing several classifiers previously used in EEG- School in San Mateo, (California), voluntarily enrolled in based BCIs for the classification of different states of anx - the study. Participants and their parents or legal guard- iety in ASD and neurotypical adolescents. We employed ians were informed extensively about the experiment classical machine learning methods, specifically support and all gave written consent. The study was approved by vector machines (SVMs) combined with FBCSP, which an Institutional Review Board composed of an educator have been successfully applied to detect a wide range from Learning Farm Educational Resources, an admin- of covert cognitive and emotional states [53], includ- istrator from The Nueva School, and a licensed mental ing mental stress detection [61–64]. Although classical health professional at The Nueva School. classifiers present several drawbacks compared to deep Participants were seated in an isolated and dimly lit learning (e.g. elaborate feature extraction and exten- room at a viewing distance of approximately 70 cm of a sive prior knowledge about the dataset [55, 65]), SVMs 16” LCD monitor with a refresh rate of 60Hz. 16-channel remain a useful benchmark against which deep learning EEG data were acquired at 125Hz using an OpenBCI sys- methods can be evaluated. tem (Ag/AgCl coated electrodes + Cyton Board; https:// For deep learning methods we selected EEGNet, a openb ci. com/) placed according to the international recently developed compact CNN for EEG-based BCIs 10–20 system (channels: ‘Fp1’, ‘Fp2’, ‘C3’, ‘C4’, ‘P7’, ‘P8’, ‘O1’, [66], as well as the Deep ConvNet and the Shallow Con- ‘O2’, ‘F7’, ‘F8’, ‘F3’, ‘F4’, ‘T7’, ‘T8’, ‘P3’, ‘P4’). The 16-electrode vNet developed by Schirrmeister et  al. [60]. Moreover, OpenBCI apparatus was selected as it maximized port- we also applied long short-term memory recurrent neu- ability, affordability, signal quality, and ease of use while ral networks (LSTM RNNs), which are a type of neural minimizing the amount of electrodes, which is ideal for net with the ability to “remember” long-term dependen- practical use of the mental stress detection system. cies far better than traditional RNNs, without the loss Participants were fitted with passive noise-canceling of short-term memory [67], and enable robust analysis headphones to isolate them from ambient noise and of temporal trends in EEG data [68]. LSTM RNNs have to interact with the stress and breath modulating inter- also shown high accuracy in emotion detection [69], with face. The audio-visual stimuli was designed in close LSTM RNN architectures performing better than CNN collaboration with Muvik Labs (https:// muvik labs. io). architectures on DEAP, a major EEG emotion analy- The stimuli featured sequential trials of stressor, guided sis dataset [55]. Hybrid deep neural networks combin- breathing, and unguided breathing sections (Fig.  1). The ing both LSTM RNN and CNN architectures have also stimuli were procedurally generated by Muvik Labs’ Aug- TM shown promising results on DEAP [70, 71]. Building mented Sound engine to ensure timing precision and upon these recent advancements, we implemented an effectiveness through evidence-backed breathing inter - LSTM RNN and a hybrid long short-term memory fully ventions driven by principles of psychoacoustics and convolutional network (LSTM-FCN) [72] to classify behavioral psychology [73]. states of mental stress from EEG. Prior to the main procedure, participants were asked The primary purpose of the present study is to assess to complete the trait anxiety component of Spielberger’s the feasibility of real-time anxiety detection based on State-Trait Anxiety Inventory for Children (STAI-C) EEG signals and the identification of a robust classifier [74], a well-validated state and trait anxiety screen used for prospective use in a pBCI able to identify the opti- for typically developing youth that can also be accurately mal breathing patterns and alleviate anxiety in students used to assess trait anxiety in children and adolescents with and without ASD. To our knowledge, this is the first with autism [75]. study to examine the efficacy of deep learning-based EEG anxiety classifiers in comparison to classical methods. In 2.2 Stress induction and alleviation addition, ours is the first attempt of EEG-based anxiety Following an initial EEG baseline recording for 120 sec classification for both adolescents with autism and neu - (‘Baseline’), participants performed a 25  min session rotypical adolescents. featuring stress induction and breath modulation tasks consisting of four main blocks. Each block began with a 2 Methods stressor featuring an augmented arithmetic number task, 2.1 P articipants and data acquisition intensified by bright contrasting colors displaying num - Eight students (1 female M: 15.13 SD: 1.45) diagnosed bers appearing sequentially, coupled with audible soni- with autism, designated as participants L1-L8, from fied timers mapped to rising pitches similar to Shepard TM Learning Farm Educational Resources based in Menlo tones (powered by Muvik Labs Augmented Sound ) Park, (California), and five students (1 female M: 16.6 [73], with a 90 second time constraint. SD: 0.55) with no known mental or neurological disor- Timed mental arithmetic has been extensively used ders, designated as participants T1-T5, from The Nueva to induce stress [76, 77]; for our specific mental stress Sundaresan et al. Brain Inf. (2021) 8:13 Page 4 of 12 Fig. 1 Experimental design of the procedure. Participants performed four blocks, each consisting of a mental arithmetic task followed by an anxiety self‑report, a period of rest, either guided breathing entrainment or unguided breathing, a second anxiety self‑report and lastly another rest period induction paradigm, we chose a widely used mental The data of two participants were rejected from all arithmetic task (for an overview see [63]) and simplified analyses due to unusually high impedances at the time of it to minimize the possibility of overstimulating par- recording, which was confirmed offline by visual inspec - ticipants with ASD. The mental stress induction was fol - tion: participant L1 from the ASD group, and participant lowed by a period of breathing for 200 seconds. The first T3 from the neurotypical group. Preprocessing of the and third breathing periods had participants breathe at EEG data was kept to a minimum to mimic online condi- their own pace (unguided breathing) while the second tions found in a real-time BCI scenario. and fourth breathing periods presented participants with For training sample preparation, a cropped training a custom-generated breathing entrainment system, guid- strategy was employed. The number of samples extracted ing breath airflow in and out of lungs at a relaxing pace for the different classifiers are shown in Table  1. Samples of around 6 brpm [78, 79] with both visual (i.e. growing/ with a length of 1 or 5s were extracted per participant shrinking circle outlining the air flow volume of target from the EEG recorded during the ‘Stressor’, ‘Guided respiration speed) and auditory guides (musical patterns Breathing’, ‘Unguided Breathing’, and ‘Baseline’ periods featuring nature sounds that mimic the sound of inhala- of the procedure and were assigned corresponding labels. TM tion and exhalation; Muvik Labs Augmented Sound ). Following each mental arithmetic task and breathing 2.4 Neural signal classification period, participants were prompted to rate their current We performed classification analysis on the selected EEG stress levels on a 5-point Likert scale, with 1 indicating training samples using an SVM model with FBCSP, three “very relaxed” and 5 indicating “very stressed”. CNN models, three LSTM RNN models, and a hybrid LSTM-FCN model. While all deep learning models were multiclass (‘Stressor’, ‘Baseline’, ‘Guided Breathing’ and 2.3 EEG sig nal processing and training data selection ‘Unguided Breathing’), the SVM classifiers were binary MNE [80], an open-source Python tool for EEG analysis, (‘Guided Breathing’ vs ‘Stressor’, ‘Unguided Breathing’ was employed to filter EEG data from all 16 channels. In vs ‘Stressor’, ‘Unguided Breathing’ vs ‘Guided Breathing’, preparation for classification analysis, EEG time-courses & ‘Baseline’ vs ‘Stressor’), as is conventional for the clas were high-pass filtered at 1 Hz to remove slow trends and sification of multiple classes with SVMs [82]. We opted subsequently low-pass filtered at 50Hz to remove line to avoid using calculated features as inputs in favor of noise. The routine clinical bandwidth for EEG is from an end-to-end learning method with filtered EEG signal 0.5 to 50Hz [81]. However, significant sinusoidal drift value inputs from all 16 channels. In addition, as differ - was observed on the 0.5Hz-1Hz interval and therefore ent EEG channels represent neural signals from differ - the interval was excluded in the selected bandpass filter ent areas of the brain, we elected not to combine channel range. data to preserve spatial information. Sundar esan et al. Brain Inf. (2021) 8:13 Page 5 of 12 For the FBCSP-SVM, the EEG recording was divided activation function. Following the first dense layer, we in the time domain into samples of 1s and partitioned in include another dropout layer with a dropout rate of 0.5. the frequency domain using 9 filter bank band-pass fil - The second dense layer consisted of 10 neurons and used ters from 4Hz to 40Hz prior to feature extraction, which a rectified linear unit (ReLU) as an activation function. was achieved with the common spatial pattern (CSP) The dense output layer of 4 neurons used softmax acti - algorithm, i.e. a linear map maximizing the variance dif- vation. The two-layer LSTM architecture is obtained by ference between two classes [83]. The binary SVM clas - omitting one 40 neuron LSTM layer, and the one-layer sifiers used a radial basis function kernel with a gamma LSTM architecture is obtained by omitting both 40 neu- value of 1/360 and a regularization parameter (C) of 1.6. ron LSTM layers (see Fig. 2C). We implemented a number of FBCSP-SVM variants, The EEGNet CNN architecture [66] used comprised such as multiclass SVM (baseline, stressor, guided and 8 2D convolutional filters of size (1, 64), a Depthwise unguided breathing) with polynomial or sigmoid kernels Convolution layer of size (16, 1) to learn multiple spa- and a 5 sec EEG sample length; these models were not tial filters for each temporal filter, a Separable Convolu - included in the comparison due to lower classification tion layer of size (1, 16), and a 4 neuron dense layer with performance. The FBCSP-SVMs were implemented with softmax activation (see Fig.  2D). In the LSTM-FCN [72] the sklearn library in Python and for validation samples architecture, EEG time series input is simultaneously fed were apportioned at a ratio of 80:20 for train and test into an LSTM block, composed of an 8 neuron LSTM dataset. layer and a dropout layer with rate of 0.8, and an FCN The Deep ConvNet CNN architecture [60] is composed block composed of 128 1D temporal convolutional layers of 4 convolution-max-pooling blocks. The first block, of size 8, 256 1D temporal convolutional layers of size 5, with 25 2D temporal convolutional filters of size (1, 5), 25 and 128 1D temporal convolutional layers of size 3. The 2D spatial convolutional filters of size (1, 64), and a max outputs of the LSTM and FCN blocks are then concate- pooling layer, was especially designed to process the EEG nated and passed into a 4 neuron dense output layer with input. The subsequent convolution-max-pooling blocks softmax activation (see Fig. 2E). each have a 2D convolutional layer and a max pooling All deep learning architectures were implemented with layer, with 50, 100 and 200 convolutional filters per block, the Keras machine learning library in Python, and were respectively. Finally, a 4 neuron dense layer with softmax trained over 1000 EEG epochs with a batch size of 200. activation produces the output (see Fig. 2A). The Shallow While training, we implemented the Adam optimization ConvNet CNN architecture is a modification of the Deep algorithm [86] with a learning rate of 0.001 in place of the ConvNet to mimic the transformations of FBCSP. The standard stochastic gradient descent (SGD) algorithm. Shallow ConvNet retains the first convolution-max-pool - During validation, EEG samples for deep learning were ing block of the Deep ConvNet, albeit with a larger kernel apportioned at a ratio of 70:30 to the train dataset and size of 13 for the temporal convolution layer. This block test dataset, respectively. performs similar transformations as the bandpass filter and CSP spatial filtering algorithm of the FBCSP work -3 Results flow. Following the convolution-max-pooling block, the 3.1 Behavioral results architecture contains a squaring nonlinearity function, On average, participants self-reported higher levels of an average pooling layer, and a logarithm activation func- mental stress on the 5-point scale following the stress tion [60]. A 4 neuron dense layer with softmax activation induction periods, with average stress scores of 1.54, produces the output (see Fig. 2B). 2.04, and 1.62 prior to 2nd, 3rd, and 4th stress induction We trained three original LSTM RNN models, with periods, respectively, and average scores of 3.00, 2.88, and one, two and three LSTM layers, respectively. The three- 3.12 following the same stress induction periods. A series layer LSTM RNN model consists of three LSTM lay- of Wilcoxon signed-rank tests were employed to compare ers, two dense hidden layers, and a dense output layer. the self-reported stress scores before and after each stress The first LSTM layer, containing 50 neurons, receives induction period across all participants; the tests indi- the input. The second and third LSTM layers contain cated that post-stressor stress scores were significantly 40 neurons. The number of neurons in the LSTM lay - higher than the pre-stressor scores, with Z test statistics ers was informed by the amount calculated and used by of −  3.19, −  2.49, and −  2.62, and p-values of 0.00143, Tsiouris et al. [84] and Alhagry et al. [69], and adjusted to 0.0127, and 0.00879, for the 2nd, 3rd, and 4th stress our EEG data to prevent underfitting and overfitting. Fol - induction periods, respectively. As the participants were lowing the third LSTM layer, we include a dropout layer prompted for their self-reported mental stress level fol- to reduce overfitting [85] with a dropout rate of 0.5. The lowing every stressor and breathing period, stress scores first dense layer contains 20 neurons and uses a sigmoid prior to the 1st stress induction period were not collected Sundaresan et al. Brain Inf. (2021) 8:13 Page 6 of 12 Fig. 2 Diagram of the model architectures for the A Deep ConvNet, B Shallow ConvNet, C LSTM RNN, D EEGNet and E LSTM‑FCN. Note: the first grayed layer of the LSTM RNN was only implemented for the two‑ and three ‑layer LSTM while the second grayed layer is only applicable to the three‑layer LSTM and hence the 1st stress induction period was not consid- baseline and unguided breathing conditions had the ered in the behavioral data analysis. Accompanied by the lowest level of demands imposed on participants, and precedents in the literature, these results reinforce our yet classification against the stressor condition yielded confidence that the selected experimental paradigm can both the highest and lowest accuracy, suggests that reliably induce mental stress in the participants. muscle activity did not bias the classifiers significantly. The three CNN models, Deep ConvNet, Shallow ConvNet, and EEGNet, yielded average classification 3.2 Model performance accuracies of 58.80, 62.84, and 61.18%, respectively The performance for the FBCSP-SVM classifiers of each (see Table  2). The LSTM-FCN yielded an average clas - participant are shown in Table  1. The average classifi - sification accuracy of 62.97% across all four classes. cation accuracy was highest for the binary classifica - The two-layer LSTM RNN classifier yielded an aver - tion of ‘Baseline’ vs ‘Stressor’ (87.88%) and lowest for age accuracy of 93.27% on the test data across all four ‘Unguided Breathing’ vs ‘Stressor’ (78.28%). That the classes, outperforming the 73.53% average accuracy of Sundar esan et al. Brain Inf. (2021) 8:13 Page 7 of 12 Table 1 Classification accuracies of the FBCSP ‑SVM classifiers inflate the accuracy metric, the two-layer LSTM RNN per participant and classification (‘Guided Breathing’ vs ‘Stressor’, model used here demonstrated high class-wise sensitiv- ‘Baseline’ vs ‘Stressor’, ‘Unguided Breathing’ vs ‘Guided Breathing’, ity and specificity during validation (see Fig.  3), leading and ‘Guided Breathing’ vs ‘Stressor ’) us to the conclusion that the unbalanced dataset was not a cause for concern. Participant Guided vs Baseline vs Unguided vs Unguided vs Stressor Stressor (%) Guided (%) Stressor (%) (%) 3.3 LSTM RNN performance with individual variation We were interested in investigating the relationship L2 98.75 97.92 100.00 96.30 between the classification accuracy of the two-layer L3 80.00 95.83 75.00 80.56 LSTM model (the best performing model) and pre-exist- L4 68.75 91.67 70.00 75.93 ing mental conditions. First, we wished to see if there was L5 81.25 91.67 87.50 77.78 a significant difference between model accuracy for par - L6 67.50 81.25 57.50 71.30 ticipants with autism and neurotypical participants. On L7 87.50 81.25 97.50 90.74 average, the two-layer LSTM model accuracy for a partic- L8 90.00 93.75 86.25 70.37 ipant with autism was 93.33%, while the model accuracy T1 90.00 93.75 76.25 71.30 for a neurotypical participant was 93.15%. A Mann- T2 83.75 89.58 91.25 83.33 Whitney U test was conducted to compare model accu- T4 72.50 79.17 73.75 69.44 racy for the participants with autism and neurotypical T5 65.00 70.83 71.25 74.04 participants and found no significant difference between Average 80.45 87.88 80.57 78.28 model accuracy for the two groups (p=0.566), indicat- ing that the two-layer LSTM model performed similarly regardless of whether the participant had autism. We also Table 2 Class‑ wise and overall accuracies for the Deep ConvNet, wished to understand whether an individual’s persistent Shallow ConvNet, and EEGNet CNN classifiers (trait) anxiety can influence the performance of the two- layer LSTM RNN. We employed Spearman correlation Average accuracy (%) Deep Shallow EEGNet (%) ConvNet ConvNet (%) to compare model accuracy and individual STAI-C trait (%) anxiety scores (see Table  4); higher STAI-C scores indi- cate higher trait anxiety. The analysis yielded a Spear - Stressor (%) 60.73 49.14 59.31 man’s rho of 0.0393, indicating virtually no correlation Unguided breathing (%) 59.38 56.25 60.38 between trait anxiety and the two-layer LSTM RNN Guided breathing (%) 53.76 81.07 59.99 performance. Baseline (%) 61.34 64.90 61.18 Average (%) 58.80 62.84 60.21 4 Discussion To the best of our knowledge, in this study we propose for the first time a deep learning-based classifier for the one-layer LSTM RNN and the 76.06% average accu- decoding mental stress, a complex and covert state, from racy of the three-layer classifier (see Table 3 ). scalp EEG signals in youth with ASD. Our results show It is important to note that due to the longer length of that states of mental stress can be accurately assessed in the unguided and guided breathing periods compared to adolescents with and without ASD as well as in adoles- the stressor and the baseline periods, more samples were cents with varying levels of baseline anxiety. We com- extracted from the unguided and guided breathing peri- pared classification accuracy of 4 binary FBCSP-SVM ods, creating an unbalanced dataset. Although this can models and 7 multiclass deep learning models. These lead to issues since an unbalanced dataset can artificially classifiers were employed to classify the EEG recorded Table 3 Class‑ wise and overall accuracies for the 1‑Layer, 2‑Layer, and 3‑Layer LSTM RNN classifiers and the hybrid LSTM ‑FCN classifier Average accuracy (%) 1-Layer LSTM (%) 2-Layer LSTM (%) 3-Layer LSTM (%) LSTM-FCN (%) Stressor (%) 63.95 90.82 74.75 57.19 Unguided breathing (%) 62.89 91.19 70.43 57.52 Guided breathing (%) 80.09 94.57 67.76 72.32 Baseline (%) 73.53 96.50 76.07 64.84 Average (%) 70.12 93.27 72.26 62.97 Highest classification accuracy are highlighted in bold Sundaresan et al. Brain Inf. (2021) 8:13 Page 8 of 12 Fig. 3 A 2‑Layer LSTM RNN model confusion matrix. B 2‑Layer LSTM RNN odel precision‑recall curve Table 4 2‑Layer LSTM RNN classification accuracy and trait conditions, which could be due to the rest periods impos- anxiety per participant ing the least, and the stress condition the most, demands on the participants. Interestingly, the classifier performed Participant 2-Layer LSTM RNN classification STAI-C trait accuracy (%) anxiety relatively well for ‘Unguided Breathing’ vs ‘Guided score Breathing’ classes, although these two conditions were similar in terms of stimuli and demands imposed on the L2 92.83 37 participants. Despite our binary FBCSP-SVM classifiers L3 92.83 32 reaching a satisfactory overall classification accuracy of L4 93.72 24 around 82% across all 4 condition pairs, there are several L5 93.69 30 trade-offs pertaining to the use of SVM when compared L6 92.57 33 to deep learning. Although SVMs require less optimizing L7 93.72 32 parameters, these learning models do not suffer from the L8 93.97 46 problem of local minima, and are less computationally T1 92.83 25 demanding than neural networks, they are constrained T2 93.08 37 to a small number of features [87], even when these fea- T4 93.47 42 tures are extracted by algorithms [88]. In addition, SVMs T5 93.24 28 cannot consider a robust set of EEG timepoints, ren- dering them unable to examine the EEG time domain, which is a critical dimension for analyses [88]. Contrast- from ASD and neurotypical adolescents performing a ingly, LSTMs are well able to handle temporal informa- task with periods of stress induction (‘Stressor’), resting tion, given their ability to choose to remember or discard state (‘Baseline’), guided breathing (‘Guided Breathing’) information depending on contextual information. None- and unguided breathing (‘Unguided Breathing’). The best theless, due to their low computational complexity, SVMs classification accuracy was achieved with the multiclass remain one of the most popular types of classifiers for two-layer LSTM at 93.27%. EEG-based BCI, in particular for online scenarios. Not- The 4 binary FBCSP-SVM classifiers performed as fol - withstanding that adaptive implementations of SVM have lows: 80.45% for ‘Guided Breathing’ vs ‘Stressor’, 87.88% been found to be superior to their static counterparts, for ‘Baseline’ vs ‘Stressor’, 80.57% for ‘Unguided Breath- they often require fully retraining the classifier with new ing’ vs ‘Guided Breathing’, and 78.28% for ‘Unguided incoming data, resulting in a much higher computa- Breathing’ vs ‘Stressor’. The FBCSP-SVM performed best tional complexity and thus a lack of online applicability when classifying between the pre-task onset resting state [53]. Conversely, with deep learning methods adaptation epoch (‘Baseline’) and the stress induction (‘Stressor’) can be achieved by retraining the input layer with new Sundar esan et al. Brain Inf. (2021) 8:13 Page 9 of 12 incoming data. LSTMs in particular are inherently adap- as a proxy for anxiety. Second, our experiment was tive and thus well suited for real-time scenarios, as their designed to induce anxiety as efficiently as possible and predictions are conditioned by past input. In addition, thus minimize the time under stress to avoid any undue unlike SVMs, deep learning networks can automatically strain on the participants. Conversely, more time was adjust and optimize their parameters, essentially alleviat- required for relaxation to set in and the breathing rate to ing the need for feature extraction and requiring far less normalize; as a result the time for the mental arithmetic processing and prior knowledge regarding the original task and the guided or unguided breathing differed. Thus, EEG dataset [55, 65]. Lastly, while some multiclass SVMs learning models were trained on an unbalanced dataset, have been found to outperform neural networks [89], our with more ‘Unguided Breathing’ and ‘Guided Breath- attempts with multiclass FBCSP-SVMs produced incon- ing’ EEG samples than ‘Stressor’ and ‘Baseline’ samples, sistent results with accuracies ranging between chance- with the potential of artificially inflating model accuracy. level and 90%. However, this is unlikely to be a concern for the two-layer With regard to the multiclass deep learning models, the LSTM RNN model, which exhibited high sensitivity and Deep ConvNet CNN performed with an overall accuracy specificity metrics across all classes. Lastly, it should be of 58.80%, the Shallow ConvNet CNN with 62.84%, the noted that a potential drawback of LSTM RNNs, and EEGNet CNN with 61.18%, the LSTM-FCN with 62.97%, of deep learning algorithms in general, is over-reliance the one-layer LSTM with 73.53%, the two-layer LSTM upon large datasets. To this regard, the same classifica - with 93.27% and the three-layer LSTM with 72.26%. The tions performed with smaller datasets including only high classification accuracies achieved with the LSTM 2 or 3 conditions led to poorer performance (data not architecture presumably is a result of its ability to learn shown). However, the experimental 2-layer LSTM accu- time dependencies within the data. Indeed, the retention racy metrics were likely not impacted by a smaller sample property is useful in mental state monitoring, as con- size and were indicative of the model’s real-world per- sidering the past activations of the EEG can drastically formance due to every trained model’s very high dem- improve the prediction of target variables and the brain onstrated sensitivity, specificity, and predictive ability; activity patterns leading up to, and associated with, spe- we found the 2-layer LSTM models were not only suc- cific cognitive states. The inherent nature of deep learn - cessful in identifying true positives across all classes but ing models, with hidden layers obscuring intermediate also when rejecting false positives in a statistically signifi - processes occurring within the models, makes it chal- cant manner in each and every one of our test subjects. lenging to definitively identify the exact cause for the Indeed, one major drawback of deep learning is the need reduction in performance with the addition of a third for large amounts of data, an issue we aim to remedy in a LSTM layer. However, it is generally understood that future study involving a much larger set of participants, stacking LSTM layers can render the model prone to both neurotypical and ASD, as well as a more diverse set overfitting [90] as well as the vanishing gradient prob - of stress induction tasks. Given that we have identified lem, in which network weights fail to update significantly a viable classifier for the monitoring of cognitive states over time and model training becomes stagnant [91, 92], related to anxiety, the goal of forthcoming studies will phenomena that could explain the lower accuracy of the be to refine and validate the two-layer LSTM RNN deep three-layer LSTM. Empirically, it has been shown that a learning model for prospective implementation in a per- second LSTM layer often provides a significant boost in sonalized pBCI. Such a system will be capable of moni- classification accuracy over a single LSTM [84, 93, 94]; toring for periods of stress and hone in on an individual’s however, the addition of a third LSTM layer or more can optimal respiration patterns by adapting the breathing have little to no effect on performance [55, 93, 94], and entrainment parameters in a closed-loop manner. in some cases additional layers can hinder model training In summary, the goal of this study was to compare and convergence, and in turn degrade performance [91]. several learning models or classifiers on their abil - Consequently, our leading hypothesis finds that the mar - ity to assess mental stress levels from EEG recordings ginal increase in model complexity between the two-layer performed on adolescent students to determine the and three-layer LSTM further complicated model train- feasibility of an EEG-based BCI capable of real-time ing while adding comparatively limited improvements, identification and the mitigation of anxiety through leading to a net loss in model performance. optimized respiration entrainment. Of the different There are some caveats to consider in the interpreta - classifiers we compared, two-layer LSTM yielded the tion of our results. First, given that anxiety varies signifi - highest classification accuracy (93.27%), opening new cantly with context and individual, and cannot therefore avenues of decoding covert mental states for BCI- be induced reliably and equivalently across participants, based neuroadaptive applications to benefit youth with we utilized mental stress induction via mental arithmetic autism. Sundaresan et al. Brain Inf. (2021) 8:13 Page 10 of 12 Acknowledgements 3. Maddox BB, White SW (2015) Comorbid social anxiety disorder in adults We would like to thank The Nueva School, Learning Farm Educational with autism spectrum disorder. J Autism Dev Disord 45:3949–3960. Resources, OpenBCI, and Muvik Labs for their kind and invaluable support. https:// doi. org/ 10. 1007/ s10803‑ 015‑ 2531‑5 The authors would also like to thank the Naturalia & Biologia Foundation for 4. van Steensel FJA, Heeman EJ (2017) Anxiety levels in children with autism financial assistance for traveling and attending meetings. spectrum disorder: a meta‑analysis. J Child Fam Stud 26:1753–1767. https:// doi. org/ 10. 1007/ s10826‑ 017‑ 0687‑7 Authors’ contributions 5. Duvekot J, van der Ende J, Verhulst FC, Greaves‑Lord K (2018) Examining AS, BP, and SC contributed equally to this work. AM, AS, AV‑ C, and VG con‑ bidirectional effects between the autism spectrum disorder (ASD) core ceived of the study idea. AM supervised the project with the help of AV‑ C. AM symptom domains and anxiety in children with ASD. J Child Psychol and VG designed and implemented the experimental paradigm. AS, BP, SC, Psychiatry 59(3):277–284. https:// doi. org/ 10. 1111/ jcpp. 12829 and AM carried out the experiment, developed the theory, and performed 6. Tomczak MT, Wójcikowski M, Pankiewicz B, Łubiński J, Majchrowicz J, the computations and analyses. SC processed the experimental data and Majchrowicz D, Walasiewicz A, Kiliński T, Szczerska M (2020) Stress moni‑ implemented the machine learning algorithms with the close help of AS and toring system for individuals with autism spectrum disorders. IEEE Access the guidance of AM. AS took the lead in writing the manuscript under the 8:228236–228244. https:// doi. org/ 10. 1109/ ACCESS. 2020. 30456 33 supervision of AM and support from BP, SC, and AV‑ C. All authors provided 7. Preece D, Howley M (2018) An approach to supporting young people critical feedback and helped shape the research, analysis and manuscript. All with autism spectrum disorder and high anxiety to re‑ engage with authors read and approved the final manuscript. formal education ‑ the impact on young people and their families. Int J Adolesc Youth 23:468–481. https:// doi. org/ 10. 1080/ 02673 843. 2018. 14336 Funding 95 The contributions of Dr. Adrien Martel and Prof. Valero‑ Cabré have been 8. Kerns CM, Kendall PC, Zickgraf H et al (2015) Not to be overshadowed or supported by research grants ANR (French National Research Agency) projet overlooked: functional impairments associated with comorbid anxiety Générique ‘OSCILOSCOPUS’ and Flag‑ era‑ JTC‑2017 CAUSALTOMICS to AV ‑ C. disorders in youth with ASD. Behav Ther 46:29–39. https:// doi. org/ 10. Dr. Martel is currently supported by a Marie Curie Fellowship (EU project 1016/j. beth. 2014. 03. 005 898813‑ CLONESA‑DLV ‑898813). 9. Clark BG, Magill‑Evans JE, Koning CJ (2015) Youth with autism spectrum disorders: self‑ and proxy‑reported quality of life and adaptive function‑ Availability of data and materials ing. Focus Autism Other Dev Disabl 30:57–64. https:// doi. org/ 10. 1177/ The informed consent forms signed by the study participants and their legal 10883 57614 522289 guardians assured anonymity and utmost confidentiality of participant data, 10. Ikeda E, Hinckson E, Krägeloh C (2014) Assessment of quality of life in so it is not possible to make the data publicly available at this time. children and youth with autism spectrum disorder: a critical review. Qual Life Res 23:1069–1085. https:// doi. org/ 10. 1007/ s11136‑ 013‑ 0591‑6 11. Cadman T, Eklund H, Howley D et al (2012) Caregiver burden as people Declarations with autism spectrum disorder and attention‑ deficit/hyperactivity disor ‑ der transition into adolescence and adulthood in the United Kingdom. J Ethics approval and consent to participate Am Acad Child Adolesc Psychiatry 51:879–888. https:// doi. org/ 10. 1016/j. The study was approved by a panel composed of an educator at Learning jaac. 2012. 06. 017 Farm Educational Resources (Menlo Park, CA), and an administrator and a 12. Smith IC, Ollendick TH, White SW (2019) Anxiety moderates the influence licensed child mental health professional at The Nueva School (San Mateo, of ASD severity on quality of life in adults with ASD. Res Autism Spectr CA). All participants and their parents or legal guardians were informed exten‑ Disord 62:39–47. https:// doi. org/ 10. 1016/j. rasd. 2019. 03. 001 sively about the study objectives and procedures, and all provided written 13. van Steensel FJA, Bögels SM, Dirksen CD (2012) Anxiety and quality of life: informed consent. clinically anxious children with and without autism spectrum disorders compared. J Clin Child Adolesc Psychol 41(6):731–738. https:// doi. org/ 10. Consent for publication 1080/ 15374 416. 2012. 698725 All authors have read the final manuscript and consented for its publication. 14. Vasa RA, Keefer A, Reaven J et al (2018) Priorities for advancing research on youth with autism spectrum disorder and co‑ occurring Competing interests anxiety. J Autism Dev Disord 48:925–934. https:// doi. org/ 10. 1007/ The authors declare that they have no competing interests. s10803‑ 017‑ 3320‑0 15. Hufnagel C, Chambres P, Bertrand PR, Dutheil F (2017) The need for Author details objective measures of stress in autism. Front Psychol 8:64. https:// doi. org/ 1 2 The Nueva School, San Mateo, CA 94033, USA. Muvik Labs, LLC, Locust 10. 3389/ fpsyg. 2017. 00064 Valley, NY 11560, USA. Center for Computer Research in Music and Acous‑ 16. Simpson K, Adams D, Wheeley E, Keen D (2020) Parent perspectives on tics, Stanford University, Stanford, CA 94305, USA. Department of Anatomy the presentation, triggers, impact, and support of anxiety in young chil‑ and Neurobiology, Laboratory of Cerebral Dynamics, Boston University School dren on the autism spectrum. J Child Fam Stud 29(2):572–582. https:// of Medicine, Boston, MA 02118, USA. Cognitive Neuroscience and Informa‑ doi. org/ 10. 1007/ s10826‑ 019‑ 01576‑5 tion Technology Research Program, Open University of Catalonia (UOC), 17. Adams D, Emerson L‑M (2020) The impact of anxiety in children on Barcelona, Spain. Causal Brain Dynamics, Plasticity and Rehabilitation Team, the autism spectrum. J Autism Dev Disord. https:// doi. org/ 10. 1007/ Frontlab, Brain and Spine Institute, ICM, CNRS UMR, Paris 7225, France. s10803‑ 020‑ 04673‑3 18. Blanck P, Perleth S, Heidenreich T et al (2018) Eec ff ts of mindfulness exer ‑ Received: 19 November 2020 Accepted: 31 May 2021 cises as stand‑alone intervention on symptoms of anxiety and depres‑ sion: Systematic review and meta‑analysis. Behav Res Ther 102:25–35. https:// doi. org/ 10. 1016/j. brat. 2017. 12. 002 19. Jayawardena R, Ranasinghe P, Ranawaka H et al (2020) Exploring the therapeutic benefits of pranayama (yogic breathing): a systematic review. References Int J Yoga 13:99–110. https:// doi. org/ 10. 4103/ ijoy. IJOY_ 37_ 19 1. American Psychiatric Association (2013) Diagnostic and statistical manual 20. Chen Y‑F, Huang X ‑ Y, Chien C‑H, Cheng J‑F (2017) The effectiveness of of mental disorders (DSM‑5). American Psychiatric Association, Virginia diaphragmatic breathing relaxation training for reducing anxiety. Per‑ 2. van Steensel FJA, Bögels SM, Perrin S (2011) Anxiety disorders in children spect Psychiatr Care 53:329–336. https:// doi. org/ 10. 1111/ ppc. 12184 and adolescents with autistic spectrum disorders: a meta‑analysis. 21. Russo MA, Santarelli DM, O’Rourke D (2017) The physiological effects Clin Child Fam Psychol Rev 14:302–317. https:// doi. org/ 10. 1007/ of slow breathing in the healthy human. Breathe (Sheff ) 13:298–309. s10567‑ 011‑ 0097‑0 https:// doi. org/ 10. 1183/ 20734 735. 009817 22. Steffen PR, Austin T, DeBarros A, Brown T (2017) The Impact of Reso ‑ nance Frequency Breathing on Measures of Heart Rate Variability, Blood Sundar esan et al. Brain Inf. (2021) 8:13 Page 11 of 12 Pressure, and Mood. Frontiers in Public Health 5:222. https:// doi. org/ 10. 43. Schultze‑Kraft M, Dähne S, Gugler M, Curio G, Blankertz B (2016) Unsuper ‑ 3389/ fpubh. 2017. 00222 vised classification of operator workload from brain signals. J Neural Eng 23. Hopper SI, Murray SL, Ferrara LR, Singleton JK (2019) Eec ff tiveness of dia‑ 13(3):036008. https:// doi. org/ 10. 1088/ 1741‑ 2560/ 13/3/ 036008 phragmatic breathing for reducing physiological and psychological stress 44. Arico P, Borghini G, Di Flumeri G et al (2016) Adaptive automation trig‑ in adults: a quantitative systematic review. JBI Database Syst Rev Imple‑ gered by EEG‑based mental workload index: a passive brain‑ computer ment Rep 17:1855–1876. https:// doi. org/ 10. 11124/ JBISR IR‑ 2017‑ 003848 interface application in realistic air traffic control environment. Front Hum 24. Xiao Y, Ma F, Lv Y et al (2015) Sustained attention is associated with error Neurosci 10:539. https:// doi. org/ 10. 3389/ fnhum. 2016. 00539 processing impairment: evidence from mental fatigue study in four‑ 45. Martel A, Dähne S, Blankertz B (2014) EEG predictors of covert vigilant choice reaction time task. PLoS One 10:e0117837. https:// doi. org/ 10. attention. J Neural Eng 11:035009. https:// doi. org/ 10. 1088/ 1741‑ 2560/ 1371/ journ al. pone. 01178 3711/3/ 035009 25. Ma X, Yue Z‑ Q, Gong Z‑ Q et al (2017) The effect of diaphragmatic 46. Van Benthem KD, Cebulski S, Herdman CM, Keillor J (2018) An EEG brain‑ breathing on attention, negative affect and stress in healthy adults. Front computer interface approach for classifying vigilance states in humans: a Psychol 8:874. https:// doi. org/ 10. 3389/ fpsyg. 2017. 00874 gamma band focus supports low misclassification rates. Int J Hum‑ Com‑ 26. Kim H‑ G, Cheon E‑ J, Bai D‑S et al (2018) Stress and heart rate variability: a put Int 34:226–237. https:// doi. org/ 10. 1080/ 10447 318. 2017. 13429 42 meta‑analysis and review of the Literature. Psychiatry Investig 15:235– 47. Liu Y, Tang Y, Yao D, Zhou X (2019) A vigilance detection method for HSR 245. https:// doi. org/ 10. 30773/ pi. 2017. 08. 17 drivers based on convolutional neural networks and wireless wearable 27. Cho H, Ryu S, Noh J, Lee J (2016) The effectiveness of daily mindful EEG collection technology. In: Transportation Research Board 98th annual breathing practices on test anxiety of students. PLoS One 11:e0164822. meeting, Washington DC, United States, 13–17 January 2019 https:// doi. org/ 10. 1371/ journ al. pone. 01648 22 48. Zhang X, Li J, Liu Y et al (2017) Design of a fatigue detection system for 28. Zaccaro A, Piarulli A, Laurino M et al (2018) How breath‑ control can high‑speed trains based on driver vigilance using a wireless wearable change your life: a systematic review on psycho‑physiological correlates EEG. Sensors. https:// doi. org/ 10. 3390/ s1703 0486 of slow breathing. Front Hum Neurosci 12:353. https:// doi. org/ 10. 3389/ 49. Roy RN, Bonnet S, Charbonnier S, Campagne A (2013) Mental fatigue fnhum. 2018. 00353 and working memory load estimation: interaction and implications for 29. Freeman FG, Mikulka PJ, Prinzel LJ, Scerbo MW (1999) Evaluation of an EEG‑based passive BCI. Conf Proc IEEE Eng Med Biol Soc 2013:6607–6610. adaptive automation system using three EEG indices with a visual track‑https:// doi. org/ 10. 1109/ EMBC. 2013. 66110 70 ing task. Biol Psychol 50:61–76. https:// doi. org/ 10. 1016/ s0301‑ 0511(99) 50. Sitaram R, Ros T, Stoeckel L et al (2017) Closed‑loop brain training: the 00002‑2 science of neurofeedback. Nat Rev Neurosci 18:86–100. https:// doi. org/ 30. Zander TO, Kothe C (2011) Towards passive brain‑ computer interfaces: 10. 1038/ nrn. 2016. 164 applying brain‑ computer interface technology to human‑machine 51. White EK, Groeneveld KM, Tittle RK et al (2017) Combined neurofeedback systems in general. J Neural Eng 8:025005. https:// doi. org/ 10. 1088/ 1741‑ and heart rate variability training for individuals with symptoms of anxi‑ 2560/8/ 2/ 025005 ety and depression: a retrospective study. NR 4:37–55. https:// doi. org/ 10. 31. Krol LR, Zander TO (2017) Passive BCI‑based neuroadaptive systems. 15540/ nr.4. 1. 37 GBCIC, Austria 52. Crivelli D, Fronda G, Balconi M (2019) Neurocognitive enhancement 32. Berka C, Levendowski DJ, Cvetinovic MM, Petrovic MM, Davis G, Lumicao effects of combined mindfulness‑neurofeedback training in sport. Neuro ‑ MN, Zivkovic VT, Popovic MV, Olmstead R (2004) Real‑time analysis of eeg science 412:83–93. https:// doi. org/ 10. 1016/j. neuro scien ce. 2019. 05. 066 indexes of alertness, cognition, and memory acquired with a wireless EEG 53. Lotte F, Bougrain L, Cichocki A et al (2018) A review of classification headset. Int J Hum‑ comput Int 17(2):151–170. https:// doi. org/ 10. 1207/ algorithms for EEG‑based brain‑ computer interfaces: a 10 year update. J s1532 7590i jhc17 02_3 Neural Eng 15:031005. https:// doi. org/ 10. 1088/ 1741‑ 2552/ aab2f2 33. McFarland DJ, Wolpaw JR (2017) EEG‑based brain‑ computer interfaces. 54. Roy Y, Banville H, Albuquerque I et al (2019) Deep learning‑based electro ‑ Curr Opin Biomed Eng 4:194–200. https:// doi. org/ 10. 1016/j. cobme. 2017. encephalography analysis: a systematic review. J Neural Eng 16:051001. 11. 004https:// doi. org/ 10. 1088/ 1741‑ 2552/ ab260c 34. Wolpaw JR, Birbaumer N, McFarland DJ et al (2002) Brain‑ computer inter‑ 55. Craik A, He Y, Contreras‑ Vidal JL (2019) Deep learning for electroen‑ faces for communication and control. Clin Neurophysiol 113:767–791. cephalogram (EEG) classification tasks: a review. J Neural Eng 16:031001. https:// doi. org/ 10. 1016/ S1388‑ 2457(02) 00057‑3https:// doi. org/ 10. 1088/ 1741‑ 2552/ ab0ab5 35. Shih JJ, Krusienski DJ, Wolpaw JR (2012) Brain‑ computer interfaces in 56. Yang H, Han J, Min K (2019) A multi‑ column CNN model for emotion medicine. Mayo Clin Proc 87:268–279. https:// doi. org/ 10. 1016/j. mayocp. recognition from EEG signals. Sensors. https:// doi. org/ 10. 3390/ s1921 4736 2011. 12. 008 57. Salama ES, El‑Khoribi RA, Shoman ME, Wahby MA, (2018) EEG‑based 36. Brunner C, Birbaumer N, Blankertz B et al (2015) BNCI Horizon 2020: emotion recognition using 3D convolutional neural networks. IJACSA. towards a roadmap for the BCI community. BCI 2:1–10. https:// doi. org/ 10. https:// doi. org/ 10. 14569/ IJACSA. 2018. 090843 1080/ 23262 63X. 2015. 10089 56 58. Wang Y, McCane B, McNaughton N et al (2019) AnxietyDecoder: an 37. Müller K‑R, Tangermann M, Dornhege G et al (2008) Machine learning EEG‑based anxiety predictor using a 3‑D convolutional neural network. for real‑time single ‑trial EEG‑analysis: from brain‑ computer interfacing to In: 2019 international joint conference on neural networks (IJCNN), Buda‑ mental state monitoring. J Neurosci Methods 167:82–90. https:// doi. org/ pest, Hungary, 14–19 July 2019, pp 1–8 10. 1016/j. jneum eth. 2007. 09. 022 59. Keng Ang K, Yang Chin Z, Zhang H, Guan C (2008) Filter bank com‑ 38. Arico P, Borghini G, Di Flumeri G et al (2017) Passive BCI in operational mon spatial pattern (FBCSP) in brain‑ computer interface. In: 2008 IEEE environments: insights, recent advances, and future trends. IEEE Trans international joint conference on neural networks (IEEE world congress Biomed Eng 64:1431–1436. https:// doi. org/ 10. 1109/ TBME. 2017. 26948 56 on computational intelligence), Hong Kong, China, 1–8 June 2008, pp 39. Clerc M, Bougrain L, Lotte F (2016) Brain‑ computer interfaces 1. Wiley‑ 2390–2397 ISTE, London 60. Schirrmeister RT, Springenberg JT, Fiederer LDJ et al (2017) Deep learning 40. Lotte F, Congedo M, Lécuyer A et al (2007) A review of classification algo ‑ with convolutional neural networks for EEG decoding and visualization. rithms for EEG‑based brain‑ computer interfaces. J Neural Eng 4:R1–R13. Hum Brain Mapp 38:5391–5420. https:// doi. org/ 10. 1002/ hbm. 23730 https:// doi. org/ 10. 1088/ 1741‑ 2560/4/ 2/ R01 61. Al‑Shargie F, Kiguchi M, Badruddin N, Dass SC, Hani AFM, Tang TB (2016) 41. Blankertz B, Curio G, Müller K‑R (2002) Classifying single trial EEG: towards Mental stress assessment using simultaneous measurement of EEG and brain computer interfacing. In: Dietterich TG, Becker S, Ghahramani Z fNIRS. Biomed Opt Express 7(10):3882–3898. https:// doi. org/ 10. 1364/ (eds) Advances in neural information processing systems 14. MIT Press, BOE.7. 003882 Cambridge, pp 157–164 62. Gaikwad P, Paithane AN (2017) Novel approach for stress recognition 42. Appriou A, Cichocki A, Lotte F (2018) Towards robust neuroadaptive HCI: using EEG signal by SVM classifier. In: 2017 international conference on exploring modern machine learning methods to estimate mental work‑ computing methodologies and communication (ICCMC), Erode, India, load from EEG signals. In: Extended abstracts of the 2018 CHI conference 18–19 July 2017, pp 967–971 on human factors in computing systems. Paper No.: LBW615. ACM, New York, pp 1–6. https:// doi. org/ 10. 1145/ 31704 27. 31886 17 Sundaresan et al. Brain Inf. (2021) 8:13 Page 12 of 12 63. Al‑Shargie F, Tang TB, Badruddin N, Kiguchi M (2018) Towards multilevel 80. Gramfort A, Luessi M, Larson E et al (2014) MNE software for processing mental stress assessment using SVM with ECOC: an EEG approach. Med MEG and EEG data. Neuroimage 86:446–460. https:// doi. org/ 10. 1016/j. Biol Eng Comput 56:125–136. https:// doi. org/ 10. 1007/ s11517‑ 017‑ 1733‑8neuro image. 2013. 10. 027 64. Saeed SMU, Anwar SM, Khalid H et al (2020) EEG based classification of 81. Vanhatalo S, Voipio J, Kaila K (2005) Full‑band EEG (FbEEG): an emerging long‑term stress using psychological labeling. Sensors. https:// doi. org/ 10. standard in electroencephalography. Clin Neurophysiol 116:1–8. https:// 3390/ s2007 1886doi. org/ 10. 1016/j. clinph. 2004. 09. 015 65. Schmidhuber J (2015) Deep learning in neural networks: an overview. 82. Duan K‑B, Keerthi SS (2005) Which is the best multiclass SVM method? An Neural Netw 61:85–117. https:// doi. org/ 10. 1016/j. neunet. 2014. 09. 003 empirical study. Multiple classifier systems. Springer, Berlin Heidelberg, pp 66. Lawhern VJ, Solon AJ, Waytowich NR et al (2018) EEGNet: a compact 278–285 convolutional neural network for EEG‑based brain‑ computer interfaces. J 83. Ramoser H, Müller‑ Gerking J, Pfurtscheller G (2000) Optimal spatial Neural Eng 15:056013. https:// doi. org/ 10. 1088/ 1741‑ 2552/ aace8c filtering of single trial EEG during imagined hand movement. IEEE Trans 67. Hochreiter S, Schmidhuber J (1997) Long short‑term memory. Neural Rehabil Eng 8:441–446. https:// doi. org/ 10. 1109/ 86. 895946 Comput 9:1735–1780. https:// doi. org/ 10. 1162/ neco. 1997.9. 8. 1735 84. Tsiouris KM, Pezoulas VC, Zervakis M, Konitsiotis S, Koutsouris DD, Fotiadis 68. Xing X, Li Z, Xu T et al (2019) SAE+LSTM: a new framework for emotion DI (2018) A long short‑term memory deep learning network for the recognition from multi‑ channel EEG. Front Neurorobot 13:37. https:// doi. prediction of epileptic seizures using EEG signals. Comput Biol Med org/ 10. 3389/ fnbot. 2019. 00037 99:24–37. https:// doi. org/ 10. 1016/j. compb iomed. 2018. 05. 019 69. Alhagry S, Fahmy AA, El‑Khoribi RA (2017) Emotion Recognition based 85. Srivastava N, Hinton G, Krizhevsky A et al (2014) Dropout: a simple way to on EEG using LSTM recurrent neural network. IJACSA. https:// doi. org/ 10. prevent neural networks from overfitting. J Mach Learn Res 15:1929–1958 14569/ IJACSA. 2017. 081046 86. Kingma DP, Ba J (2014) Adam: a method for stochastic optimization. In: 70. Yang Y, Wu Q, Qiu M et al (2018) Emotion recognition from multi‑ channel 3rd international conference for learning representations, San Diego, EEG through parallel convolutional recurrent neural network. In: 2018 2015. arXiv: 1412. 6980 international joint conference on neural networks (IJCNN), Rio de Janeiro, 87. Faust O, Hagiwara Y, Hong TJ et al (2018) Deep learning for healthcare Brazil, 8–13 July 2018, pp 1–7 applications based on physiological signals: A review. Comput Methods 71. Li Y, Huang J, Zhou H, Zhong N (2017) Human emotion recognition with Programs Biomed 161:1–13. https:// doi. org/ 10. 1016/j. cmpb. 2018. 04. 005 electroencephalographic multidimensional features by hybrid deep 88. Vahid A, Mückschel M, Stober S et al (2020) Applying deep learning to neural networks. NATO Adv Sci Inst Ser E Appl Sci 7:1060. https:// doi. org/ single‑trial EEG data provides evidence for complementary theo ‑ 10. 3390/ app71 01060 ries on action control. Commun Biol 3:112. https:// doi. org/ 10. 1038/ 72. Karim F, Majumdar S, Darabi H, Chen S (2018) LSTM fully convolutional s42003‑ 020‑ 0846‑z networks for time series classification. IEEE Access 6:1662–1669. https:// 89. Güler I, Ubeyli ED (2007) Multiclass support vector machines for EEG‑ doi. org/ 10. 1109/ ACCESS. 2017. 27799 39 signals classification. IEEE Trans Inf Technol Biomed 11:117–126. https:// 73. Borthakur D, Grace V, Batchelor P et al (2019) Fuzzy C‑means cluster ‑doi. org/ 10. 1109/ titb. 2006. 879600 ing and sonification of HRV features. In: 2019 IEEE/ACM international 90. Dai Yin, Jha, (2020) Grow and prune compact, fast, and accurate LSTMs. conference on connected health: applications, systems and engineering IEEE Trans Comput 69:441–452. https:// doi. org/ 10. 1109/ TC. 2019. 29544 95 technologies (CHASE), Arlington, VA, USA, 25–27 September 2019, pp 91. Zhang Y, Chen G, Yu D, Yao K, Khudanpur S, Glass J (2016) Highway long 53–57 short‑term memory RNNS for distant speech recognition. In: 2016 IEEE 74. Spielberger CD (1973) Manual for the state‑trait inventory for children. international conference on acoustics, speech and signal processing Palo Alto, Santa Clara (ICASSP), Shanghai, China, 20–25 March 2016, pp 5755–5759. https:// doi. 75. Simon DM, Corbett BA (2013) Examining associations between anxiety org/ 10. 1109/ ICASSP. 2016. 74727 80 and cortisol in high functioning male children with autism. J Neurodev 92. Hsu W‑N, Zhang Y, Glass J (2016) A prioritized grid long short ‑term Disord 5:32. https:// doi. org/ 10. 1186/ 1866‑ 1955‑5‑ 32 memory RNN for speech recognition. SLT 2016:467–473. https:// doi. org/ 76. Dedovic K, Renwick R, Mahani NK et al (2005) The montreal imaging 10. 1109/ SLT. 2016. 78463 05 stress task: using functional imaging to investigate the effects of perceiv‑ 93. Tseng P‑H, Urpi NA, Lebedev M, Nicolelis M (2019) Decoding movements ing and processing psychosocial stress in the human brain. J Psychiatry from cortical ensemble activity using a long short‑term memory recur ‑ Neurosci 30:319–325 rent network. Neural Comput 31(6):1085–1113. https:// doi. org/ 10. 1162/ 77. Shilton AL, Laycock R, Crewther SG (2017) The maastricht acute stress test neco_a_ 01189 (MAST ): physiological and subjective responses in anticipation, and post‑ 94. Wang P, Jiang A, Liu X, Shang J, Zhang L (2018) LSTM‑based EEG clas‑ stress. Front Psychol 8:567. https:// doi. org/ 10. 3389/ fpsyg. 2017. 00567 sification in motor imagery tasks. IEEE Trans Neural Syst Rehabil Eng 78. Szulczewski MT (2019) Training of paced breathing at 0.1 Hz improves 26(11):2086–2095. https:// doi. org/ 10. 1109/ TNSRE. 2018. 28761 29 CO2 homeostasis and relaxation during a paced breathing task. PLoS One 14. https:// doi. org/ 10. 1371/ journ al. pone. 02185 50 Publisher’s Note 79. Lin IM, Tai LY, Fan SY (2014) Breathing at a rate of 5.5 breaths per minute Springer Nature remains neutral with regard to jurisdictional claims in pub‑ with equal inhalation‑to ‑ exhalation ratio increases heart rate variability. lished maps and institutional affiliations. Int J Psychophysiol 91:206–211. https:// doi. org/ 10. 1016/j. ijpsy cho. 2013. 12. 006 http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Brain Informatics Springer Journals

Evaluating deep learning EEG-based mental stress classification in adolescents with autism for breathing entrainment BCI

Download PDF
12 pages

Loading next page...
 
/lp/springer-journals/evaluating-deep-learning-eeg-based-mental-stress-classification-in-xu9uZIMPpF

References (92)

Publisher
Springer Journals
Copyright
Copyright © The Author(s) 2021
ISSN
2198-4018
eISSN
2198-4026
DOI
10.1186/s40708-021-00133-5
Publisher site
See Article on Publisher Site

Abstract

Mental stress is a major individual and societal burden and one of the main contributing factors that lead to patholo‑ gies such as depression, anxiety disorders, heart attacks, and strokes. Given that anxiety disorders are one of the most common comorbidities in youth with autism spectrum disorder (ASD), this population is particularly vulnerable to mental stress, severely limiting overall quality of life. To prevent this, early stress quantification with machine learning (ML) and effective anxiety mitigation with non‑pharmacological interventions are essential. This study aims to investi‑ gate the feasibility of exploiting electroencephalography (EEG) signals for stress assessment by comparing several ML classifiers, namely support vector machine (SVM) and deep learning methods. We trained a total of eleven subject ‑ dependent models‑four with conventional brain‑ computer interface (BCI) methods and seven with deep learning approaches‑ on the EEG of neurotypical (n=5) and ASD (n=8) participants performing alternating blocks of mental arithmetic stress induction, guided and unguided breathing. Our results show that a multiclass two‑layer LSTM RNN deep learning classifier is capable of identifying mental stress from ongoing EEG with an overall accuracy of 93.27%. Our study is the first to successfully apply an LSTM RNN classifier to identify stress states from EEG in both ASD and neurotypical adolescents, and offers promise for an EEG‑based BCI for the real‑time assessment and mitigation of mental stress through a closed‑loop adaptation of respiration entrainment. Keywords: Mental stress, Autism, EEG, Deep learning, Breathing entrainment 1 Introduction of stress even in the absence of these stressors. The Individuals with autism spectrum disorder (ASD) often comorbidity of anxiety disorders and ASD in children demonstrate deficits in social communication skills and and adolescents has been studied extensively with 40% restricted or stereotyped behaviors and interests [1]. This to 85% of individuals with ASD aged 6 to 18 having at causes those with ASD to experience states of cogni- least one form of anxiety [3–5]. Unfortunately, individu- tive and emotional overload, leading to increased stress als with ASD are uniquely vulnerable to the deleterious and ultimately anxiety symptoms [2]. Although there effects of stress because of their hyper- or hyporeactiv - is significant overlap between stress and anxiety, stress ity to sensory inputs, as well as difficulties with accurate is best understood as the physiological and psychologi- stress detection and coping with stressful situations [6]. cal response towards stressors; anxiety is the persistence Given the frequency in which anxiety co-occurs in ASD, in conjunction with the hurdles in education, long-term functional impairments, reduction in quality of life, and *Correspondence: adrien.martel@icm‑institute.org increased caregiver burden [7–13], a more comprehen- Cognitive Neuroscience and Information Technology Research Program, sive understanding of comorbidities in ASD as well as Open University of Catalonia (UOC), Barcelona, Spain personalized intervention methods to relieve clinical Full list of author information is available at the end of the article © The Author(s) 2021. This article is licensed under a Creative Commons Attribution 4.0 International License, which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if changes were made. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http:// creat iveco mmons. org/ licen ses/ by/4. 0/. Sundaresan et al. Brain Inf. (2021) 8:13 Page 2 of 12 symptoms of the disease and improve emotional and and passive brain-computer interfaces (pBCIs) aim at physical well-being for individuals with ASD is needed. intelligent forms of adaptation in response to cognitive Incidentally, anxiety and the design of appropriate inter- state assessments [30, 31]. The field of EEG-based BCIs vention methods have been identified by the autism com - has blossomed in recent years, largely on account of munity and clinicians as a key priority with researchers EEG’s high temporal resolution, non-invasiveness, rela- emphasizing the need for more precise measures of anxi- tively low cost, and novel advances in the effectiveness ety [14]. Moreover, the lack of objective and continuous and usability of acquisition systems [32, 33]. While BCIs measurements of stress is particularly detrimental for have historically been employed in the context of assis- a population already affected by an inability to express tive technologies for severely impaired individuals [34, inner experiences and calls for novel methods to identify 35], pBCIs have mainly been aimed at developing adap- individualized stress markers in real-time [15]. Among tive automation for real-world applications [36–38]. The the triggers identified, such as challenging sensory expe - central challenge of EEG-based pBCIs is to account for riences or social demands, anxiety related to academic the high inter- and intra-individual variability of neuro- expectations is thought to have the greatest impact on physiological signals exhibited under particular cognitive school performance for ASD children and adolescents states [39]. However, by averaging over a large enough [16, 17]. number of samples it is possible to distill sufficiently spe - Concurrently, a growing number of studies have dem- cific brain activity patterns and train a machine learning onstrated the efficacy of stand-alone meditation, relaxa - classifier to learn to discern these patterns in real-time tion and breathing practices for improving well-being, [40, 41]. This approach has already been successfully mental health and managing stress [18–20]. Although applied to monitor several cognitive states such as work- the underlying mechanisms are not yet fully understood, load [42–44], vigilance [45–47], and fatigue [48, 49]. breathing practices such as ‘anatomically optimized res- Neurofeedback involves monitoring a user’s mental piration’, i.e. controlled, slow diaphragmatic breathing state with EEG and providing feedback through a vari- through the nose in the range of 6-10 breaths per min- ety of modalities (visual, audio, tactile, etc.) to modulate ute (brpm) or resonance frequency breathing, 4.5 to 6.5 particular biomarkers [50]. In conjunction with breath- brpm for adults and 6.5 to 9.5 brpm for children, have control, neurofeedback training has been shown to be been found to procure significant physiological benefits a promising mitigatory tool for anxiety. For example, [21, 22], reduce physiological and psychological stress White et  al. [51] demonstrated that breathing-based [23] and even improve sustained attention performance visual neurofeedback reduces symptoms in patients with [24]. Prior studies have shown that breath-control can anxiety and depression, while acoustically mediated deep address physiological correlates of anxiety, including breathing neurofeedback was shown by Crivelli et al. [52] heart rate variability [22, 25], a well-validated quantita- to diminish long-term stress and anxiety levels in young tive stress indicator [26]. Notably, breath-control has adults. The first step towards an EEG-based BCI able to been found to significantly decrease test anxiety in stu - monitor anxiety levels, identify an individual’s optimal dents in an educational setting [27]. Moreover, a recent breathing patterns, and adapt breathing entrainment review by Zaccaro et  al. [28] found that controlled slow parameters in real-time, is to determine whether mental breathing (<10 brpm) had a significant impact on auto - stress can be classified on the basis of ongoing EEG data. nomic nervous system activity, especially in the theta Classification algorithms are key elements of any EEG- (3–7 Hz), alpha (8–14 Hz) and beta (15–30Hz) bands based BCI’s ability to recognize users’ EEG patterns and of the electroencephalogram (EEG), linked to improved associated cognitive states. Among the large diversity of cognitive performance during attentional and executive existing architectures and types of classifiers (for reviews functions [29]. Although these findings, taken together, see [53] and [54]), deep learning methods have recently speak to the promise of using controlled slow breathing emerged as methods of analysis that can consider neu- as a simple, low-cost and non-pharmacologic interven- rophysiological data in its entirety, including the time tion [23] to mitigate anxiety, optimized efficacy hinges on domain [54]. Convolutional neural networks are the assessing an individual’s current level of stress and ideal most widely used deep learning algorithms in EEG analy- respiration parameters in real-time. sis [55], and have been shown to be effective in emotion Although cognitive or affective states such as stress are detection [56, 57] and anxiety classification [58] in par - not directly observable externally nor reliably measurable ticular. Further, deep learning with convolutional neural through behavioral measures or subjective reports, devel- networks (CNNs) have recently been shown to outper- opments in EEG-based brain-computer interfaces (BCIs) form the widely used filter bank common spatial pattern have increasingly permitted the continuous and real-time (FBCSP) algorithm [59] by extracting increasingly more monitoring of mental states. Neuroadaptive technologies complex features of the data [60]. Accordingly, we aimed Sundar esan et al. Brain Inf. (2021) 8:13 Page 3 of 12 at comparing several classifiers previously used in EEG- School in San Mateo, (California), voluntarily enrolled in based BCIs for the classification of different states of anx - the study. Participants and their parents or legal guard- iety in ASD and neurotypical adolescents. We employed ians were informed extensively about the experiment classical machine learning methods, specifically support and all gave written consent. The study was approved by vector machines (SVMs) combined with FBCSP, which an Institutional Review Board composed of an educator have been successfully applied to detect a wide range from Learning Farm Educational Resources, an admin- of covert cognitive and emotional states [53], includ- istrator from The Nueva School, and a licensed mental ing mental stress detection [61–64]. Although classical health professional at The Nueva School. classifiers present several drawbacks compared to deep Participants were seated in an isolated and dimly lit learning (e.g. elaborate feature extraction and exten- room at a viewing distance of approximately 70 cm of a sive prior knowledge about the dataset [55, 65]), SVMs 16” LCD monitor with a refresh rate of 60Hz. 16-channel remain a useful benchmark against which deep learning EEG data were acquired at 125Hz using an OpenBCI sys- methods can be evaluated. tem (Ag/AgCl coated electrodes + Cyton Board; https:// For deep learning methods we selected EEGNet, a openb ci. com/) placed according to the international recently developed compact CNN for EEG-based BCIs 10–20 system (channels: ‘Fp1’, ‘Fp2’, ‘C3’, ‘C4’, ‘P7’, ‘P8’, ‘O1’, [66], as well as the Deep ConvNet and the Shallow Con- ‘O2’, ‘F7’, ‘F8’, ‘F3’, ‘F4’, ‘T7’, ‘T8’, ‘P3’, ‘P4’). The 16-electrode vNet developed by Schirrmeister et  al. [60]. Moreover, OpenBCI apparatus was selected as it maximized port- we also applied long short-term memory recurrent neu- ability, affordability, signal quality, and ease of use while ral networks (LSTM RNNs), which are a type of neural minimizing the amount of electrodes, which is ideal for net with the ability to “remember” long-term dependen- practical use of the mental stress detection system. cies far better than traditional RNNs, without the loss Participants were fitted with passive noise-canceling of short-term memory [67], and enable robust analysis headphones to isolate them from ambient noise and of temporal trends in EEG data [68]. LSTM RNNs have to interact with the stress and breath modulating inter- also shown high accuracy in emotion detection [69], with face. The audio-visual stimuli was designed in close LSTM RNN architectures performing better than CNN collaboration with Muvik Labs (https:// muvik labs. io). architectures on DEAP, a major EEG emotion analy- The stimuli featured sequential trials of stressor, guided sis dataset [55]. Hybrid deep neural networks combin- breathing, and unguided breathing sections (Fig.  1). The ing both LSTM RNN and CNN architectures have also stimuli were procedurally generated by Muvik Labs’ Aug- TM shown promising results on DEAP [70, 71]. Building mented Sound engine to ensure timing precision and upon these recent advancements, we implemented an effectiveness through evidence-backed breathing inter - LSTM RNN and a hybrid long short-term memory fully ventions driven by principles of psychoacoustics and convolutional network (LSTM-FCN) [72] to classify behavioral psychology [73]. states of mental stress from EEG. Prior to the main procedure, participants were asked The primary purpose of the present study is to assess to complete the trait anxiety component of Spielberger’s the feasibility of real-time anxiety detection based on State-Trait Anxiety Inventory for Children (STAI-C) EEG signals and the identification of a robust classifier [74], a well-validated state and trait anxiety screen used for prospective use in a pBCI able to identify the opti- for typically developing youth that can also be accurately mal breathing patterns and alleviate anxiety in students used to assess trait anxiety in children and adolescents with and without ASD. To our knowledge, this is the first with autism [75]. study to examine the efficacy of deep learning-based EEG anxiety classifiers in comparison to classical methods. In 2.2 Stress induction and alleviation addition, ours is the first attempt of EEG-based anxiety Following an initial EEG baseline recording for 120 sec classification for both adolescents with autism and neu - (‘Baseline’), participants performed a 25  min session rotypical adolescents. featuring stress induction and breath modulation tasks consisting of four main blocks. Each block began with a 2 Methods stressor featuring an augmented arithmetic number task, 2.1 P articipants and data acquisition intensified by bright contrasting colors displaying num - Eight students (1 female M: 15.13 SD: 1.45) diagnosed bers appearing sequentially, coupled with audible soni- with autism, designated as participants L1-L8, from fied timers mapped to rising pitches similar to Shepard TM Learning Farm Educational Resources based in Menlo tones (powered by Muvik Labs Augmented Sound ) Park, (California), and five students (1 female M: 16.6 [73], with a 90 second time constraint. SD: 0.55) with no known mental or neurological disor- Timed mental arithmetic has been extensively used ders, designated as participants T1-T5, from The Nueva to induce stress [76, 77]; for our specific mental stress Sundaresan et al. Brain Inf. (2021) 8:13 Page 4 of 12 Fig. 1 Experimental design of the procedure. Participants performed four blocks, each consisting of a mental arithmetic task followed by an anxiety self‑report, a period of rest, either guided breathing entrainment or unguided breathing, a second anxiety self‑report and lastly another rest period induction paradigm, we chose a widely used mental The data of two participants were rejected from all arithmetic task (for an overview see [63]) and simplified analyses due to unusually high impedances at the time of it to minimize the possibility of overstimulating par- recording, which was confirmed offline by visual inspec - ticipants with ASD. The mental stress induction was fol - tion: participant L1 from the ASD group, and participant lowed by a period of breathing for 200 seconds. The first T3 from the neurotypical group. Preprocessing of the and third breathing periods had participants breathe at EEG data was kept to a minimum to mimic online condi- their own pace (unguided breathing) while the second tions found in a real-time BCI scenario. and fourth breathing periods presented participants with For training sample preparation, a cropped training a custom-generated breathing entrainment system, guid- strategy was employed. The number of samples extracted ing breath airflow in and out of lungs at a relaxing pace for the different classifiers are shown in Table  1. Samples of around 6 brpm [78, 79] with both visual (i.e. growing/ with a length of 1 or 5s were extracted per participant shrinking circle outlining the air flow volume of target from the EEG recorded during the ‘Stressor’, ‘Guided respiration speed) and auditory guides (musical patterns Breathing’, ‘Unguided Breathing’, and ‘Baseline’ periods featuring nature sounds that mimic the sound of inhala- of the procedure and were assigned corresponding labels. TM tion and exhalation; Muvik Labs Augmented Sound ). Following each mental arithmetic task and breathing 2.4 Neural signal classification period, participants were prompted to rate their current We performed classification analysis on the selected EEG stress levels on a 5-point Likert scale, with 1 indicating training samples using an SVM model with FBCSP, three “very relaxed” and 5 indicating “very stressed”. CNN models, three LSTM RNN models, and a hybrid LSTM-FCN model. While all deep learning models were multiclass (‘Stressor’, ‘Baseline’, ‘Guided Breathing’ and 2.3 EEG sig nal processing and training data selection ‘Unguided Breathing’), the SVM classifiers were binary MNE [80], an open-source Python tool for EEG analysis, (‘Guided Breathing’ vs ‘Stressor’, ‘Unguided Breathing’ was employed to filter EEG data from all 16 channels. In vs ‘Stressor’, ‘Unguided Breathing’ vs ‘Guided Breathing’, preparation for classification analysis, EEG time-courses & ‘Baseline’ vs ‘Stressor’), as is conventional for the clas were high-pass filtered at 1 Hz to remove slow trends and sification of multiple classes with SVMs [82]. We opted subsequently low-pass filtered at 50Hz to remove line to avoid using calculated features as inputs in favor of noise. The routine clinical bandwidth for EEG is from an end-to-end learning method with filtered EEG signal 0.5 to 50Hz [81]. However, significant sinusoidal drift value inputs from all 16 channels. In addition, as differ - was observed on the 0.5Hz-1Hz interval and therefore ent EEG channels represent neural signals from differ - the interval was excluded in the selected bandpass filter ent areas of the brain, we elected not to combine channel range. data to preserve spatial information. Sundar esan et al. Brain Inf. (2021) 8:13 Page 5 of 12 For the FBCSP-SVM, the EEG recording was divided activation function. Following the first dense layer, we in the time domain into samples of 1s and partitioned in include another dropout layer with a dropout rate of 0.5. the frequency domain using 9 filter bank band-pass fil - The second dense layer consisted of 10 neurons and used ters from 4Hz to 40Hz prior to feature extraction, which a rectified linear unit (ReLU) as an activation function. was achieved with the common spatial pattern (CSP) The dense output layer of 4 neurons used softmax acti - algorithm, i.e. a linear map maximizing the variance dif- vation. The two-layer LSTM architecture is obtained by ference between two classes [83]. The binary SVM clas - omitting one 40 neuron LSTM layer, and the one-layer sifiers used a radial basis function kernel with a gamma LSTM architecture is obtained by omitting both 40 neu- value of 1/360 and a regularization parameter (C) of 1.6. ron LSTM layers (see Fig. 2C). We implemented a number of FBCSP-SVM variants, The EEGNet CNN architecture [66] used comprised such as multiclass SVM (baseline, stressor, guided and 8 2D convolutional filters of size (1, 64), a Depthwise unguided breathing) with polynomial or sigmoid kernels Convolution layer of size (16, 1) to learn multiple spa- and a 5 sec EEG sample length; these models were not tial filters for each temporal filter, a Separable Convolu - included in the comparison due to lower classification tion layer of size (1, 16), and a 4 neuron dense layer with performance. The FBCSP-SVMs were implemented with softmax activation (see Fig.  2D). In the LSTM-FCN [72] the sklearn library in Python and for validation samples architecture, EEG time series input is simultaneously fed were apportioned at a ratio of 80:20 for train and test into an LSTM block, composed of an 8 neuron LSTM dataset. layer and a dropout layer with rate of 0.8, and an FCN The Deep ConvNet CNN architecture [60] is composed block composed of 128 1D temporal convolutional layers of 4 convolution-max-pooling blocks. The first block, of size 8, 256 1D temporal convolutional layers of size 5, with 25 2D temporal convolutional filters of size (1, 5), 25 and 128 1D temporal convolutional layers of size 3. The 2D spatial convolutional filters of size (1, 64), and a max outputs of the LSTM and FCN blocks are then concate- pooling layer, was especially designed to process the EEG nated and passed into a 4 neuron dense output layer with input. The subsequent convolution-max-pooling blocks softmax activation (see Fig. 2E). each have a 2D convolutional layer and a max pooling All deep learning architectures were implemented with layer, with 50, 100 and 200 convolutional filters per block, the Keras machine learning library in Python, and were respectively. Finally, a 4 neuron dense layer with softmax trained over 1000 EEG epochs with a batch size of 200. activation produces the output (see Fig. 2A). The Shallow While training, we implemented the Adam optimization ConvNet CNN architecture is a modification of the Deep algorithm [86] with a learning rate of 0.001 in place of the ConvNet to mimic the transformations of FBCSP. The standard stochastic gradient descent (SGD) algorithm. Shallow ConvNet retains the first convolution-max-pool - During validation, EEG samples for deep learning were ing block of the Deep ConvNet, albeit with a larger kernel apportioned at a ratio of 70:30 to the train dataset and size of 13 for the temporal convolution layer. This block test dataset, respectively. performs similar transformations as the bandpass filter and CSP spatial filtering algorithm of the FBCSP work -3 Results flow. Following the convolution-max-pooling block, the 3.1 Behavioral results architecture contains a squaring nonlinearity function, On average, participants self-reported higher levels of an average pooling layer, and a logarithm activation func- mental stress on the 5-point scale following the stress tion [60]. A 4 neuron dense layer with softmax activation induction periods, with average stress scores of 1.54, produces the output (see Fig. 2B). 2.04, and 1.62 prior to 2nd, 3rd, and 4th stress induction We trained three original LSTM RNN models, with periods, respectively, and average scores of 3.00, 2.88, and one, two and three LSTM layers, respectively. The three- 3.12 following the same stress induction periods. A series layer LSTM RNN model consists of three LSTM lay- of Wilcoxon signed-rank tests were employed to compare ers, two dense hidden layers, and a dense output layer. the self-reported stress scores before and after each stress The first LSTM layer, containing 50 neurons, receives induction period across all participants; the tests indi- the input. The second and third LSTM layers contain cated that post-stressor stress scores were significantly 40 neurons. The number of neurons in the LSTM lay - higher than the pre-stressor scores, with Z test statistics ers was informed by the amount calculated and used by of −  3.19, −  2.49, and −  2.62, and p-values of 0.00143, Tsiouris et al. [84] and Alhagry et al. [69], and adjusted to 0.0127, and 0.00879, for the 2nd, 3rd, and 4th stress our EEG data to prevent underfitting and overfitting. Fol - induction periods, respectively. As the participants were lowing the third LSTM layer, we include a dropout layer prompted for their self-reported mental stress level fol- to reduce overfitting [85] with a dropout rate of 0.5. The lowing every stressor and breathing period, stress scores first dense layer contains 20 neurons and uses a sigmoid prior to the 1st stress induction period were not collected Sundaresan et al. Brain Inf. (2021) 8:13 Page 6 of 12 Fig. 2 Diagram of the model architectures for the A Deep ConvNet, B Shallow ConvNet, C LSTM RNN, D EEGNet and E LSTM‑FCN. Note: the first grayed layer of the LSTM RNN was only implemented for the two‑ and three ‑layer LSTM while the second grayed layer is only applicable to the three‑layer LSTM and hence the 1st stress induction period was not consid- baseline and unguided breathing conditions had the ered in the behavioral data analysis. Accompanied by the lowest level of demands imposed on participants, and precedents in the literature, these results reinforce our yet classification against the stressor condition yielded confidence that the selected experimental paradigm can both the highest and lowest accuracy, suggests that reliably induce mental stress in the participants. muscle activity did not bias the classifiers significantly. The three CNN models, Deep ConvNet, Shallow ConvNet, and EEGNet, yielded average classification 3.2 Model performance accuracies of 58.80, 62.84, and 61.18%, respectively The performance for the FBCSP-SVM classifiers of each (see Table  2). The LSTM-FCN yielded an average clas - participant are shown in Table  1. The average classifi - sification accuracy of 62.97% across all four classes. cation accuracy was highest for the binary classifica - The two-layer LSTM RNN classifier yielded an aver - tion of ‘Baseline’ vs ‘Stressor’ (87.88%) and lowest for age accuracy of 93.27% on the test data across all four ‘Unguided Breathing’ vs ‘Stressor’ (78.28%). That the classes, outperforming the 73.53% average accuracy of Sundar esan et al. Brain Inf. (2021) 8:13 Page 7 of 12 Table 1 Classification accuracies of the FBCSP ‑SVM classifiers inflate the accuracy metric, the two-layer LSTM RNN per participant and classification (‘Guided Breathing’ vs ‘Stressor’, model used here demonstrated high class-wise sensitiv- ‘Baseline’ vs ‘Stressor’, ‘Unguided Breathing’ vs ‘Guided Breathing’, ity and specificity during validation (see Fig.  3), leading and ‘Guided Breathing’ vs ‘Stressor ’) us to the conclusion that the unbalanced dataset was not a cause for concern. Participant Guided vs Baseline vs Unguided vs Unguided vs Stressor Stressor (%) Guided (%) Stressor (%) (%) 3.3 LSTM RNN performance with individual variation We were interested in investigating the relationship L2 98.75 97.92 100.00 96.30 between the classification accuracy of the two-layer L3 80.00 95.83 75.00 80.56 LSTM model (the best performing model) and pre-exist- L4 68.75 91.67 70.00 75.93 ing mental conditions. First, we wished to see if there was L5 81.25 91.67 87.50 77.78 a significant difference between model accuracy for par - L6 67.50 81.25 57.50 71.30 ticipants with autism and neurotypical participants. On L7 87.50 81.25 97.50 90.74 average, the two-layer LSTM model accuracy for a partic- L8 90.00 93.75 86.25 70.37 ipant with autism was 93.33%, while the model accuracy T1 90.00 93.75 76.25 71.30 for a neurotypical participant was 93.15%. A Mann- T2 83.75 89.58 91.25 83.33 Whitney U test was conducted to compare model accu- T4 72.50 79.17 73.75 69.44 racy for the participants with autism and neurotypical T5 65.00 70.83 71.25 74.04 participants and found no significant difference between Average 80.45 87.88 80.57 78.28 model accuracy for the two groups (p=0.566), indicat- ing that the two-layer LSTM model performed similarly regardless of whether the participant had autism. We also Table 2 Class‑ wise and overall accuracies for the Deep ConvNet, wished to understand whether an individual’s persistent Shallow ConvNet, and EEGNet CNN classifiers (trait) anxiety can influence the performance of the two- layer LSTM RNN. We employed Spearman correlation Average accuracy (%) Deep Shallow EEGNet (%) ConvNet ConvNet (%) to compare model accuracy and individual STAI-C trait (%) anxiety scores (see Table  4); higher STAI-C scores indi- cate higher trait anxiety. The analysis yielded a Spear - Stressor (%) 60.73 49.14 59.31 man’s rho of 0.0393, indicating virtually no correlation Unguided breathing (%) 59.38 56.25 60.38 between trait anxiety and the two-layer LSTM RNN Guided breathing (%) 53.76 81.07 59.99 performance. Baseline (%) 61.34 64.90 61.18 Average (%) 58.80 62.84 60.21 4 Discussion To the best of our knowledge, in this study we propose for the first time a deep learning-based classifier for the one-layer LSTM RNN and the 76.06% average accu- decoding mental stress, a complex and covert state, from racy of the three-layer classifier (see Table 3 ). scalp EEG signals in youth with ASD. Our results show It is important to note that due to the longer length of that states of mental stress can be accurately assessed in the unguided and guided breathing periods compared to adolescents with and without ASD as well as in adoles- the stressor and the baseline periods, more samples were cents with varying levels of baseline anxiety. We com- extracted from the unguided and guided breathing peri- pared classification accuracy of 4 binary FBCSP-SVM ods, creating an unbalanced dataset. Although this can models and 7 multiclass deep learning models. These lead to issues since an unbalanced dataset can artificially classifiers were employed to classify the EEG recorded Table 3 Class‑ wise and overall accuracies for the 1‑Layer, 2‑Layer, and 3‑Layer LSTM RNN classifiers and the hybrid LSTM ‑FCN classifier Average accuracy (%) 1-Layer LSTM (%) 2-Layer LSTM (%) 3-Layer LSTM (%) LSTM-FCN (%) Stressor (%) 63.95 90.82 74.75 57.19 Unguided breathing (%) 62.89 91.19 70.43 57.52 Guided breathing (%) 80.09 94.57 67.76 72.32 Baseline (%) 73.53 96.50 76.07 64.84 Average (%) 70.12 93.27 72.26 62.97 Highest classification accuracy are highlighted in bold Sundaresan et al. Brain Inf. (2021) 8:13 Page 8 of 12 Fig. 3 A 2‑Layer LSTM RNN model confusion matrix. B 2‑Layer LSTM RNN odel precision‑recall curve Table 4 2‑Layer LSTM RNN classification accuracy and trait conditions, which could be due to the rest periods impos- anxiety per participant ing the least, and the stress condition the most, demands on the participants. Interestingly, the classifier performed Participant 2-Layer LSTM RNN classification STAI-C trait accuracy (%) anxiety relatively well for ‘Unguided Breathing’ vs ‘Guided score Breathing’ classes, although these two conditions were similar in terms of stimuli and demands imposed on the L2 92.83 37 participants. Despite our binary FBCSP-SVM classifiers L3 92.83 32 reaching a satisfactory overall classification accuracy of L4 93.72 24 around 82% across all 4 condition pairs, there are several L5 93.69 30 trade-offs pertaining to the use of SVM when compared L6 92.57 33 to deep learning. Although SVMs require less optimizing L7 93.72 32 parameters, these learning models do not suffer from the L8 93.97 46 problem of local minima, and are less computationally T1 92.83 25 demanding than neural networks, they are constrained T2 93.08 37 to a small number of features [87], even when these fea- T4 93.47 42 tures are extracted by algorithms [88]. In addition, SVMs T5 93.24 28 cannot consider a robust set of EEG timepoints, ren- dering them unable to examine the EEG time domain, which is a critical dimension for analyses [88]. Contrast- from ASD and neurotypical adolescents performing a ingly, LSTMs are well able to handle temporal informa- task with periods of stress induction (‘Stressor’), resting tion, given their ability to choose to remember or discard state (‘Baseline’), guided breathing (‘Guided Breathing’) information depending on contextual information. None- and unguided breathing (‘Unguided Breathing’). The best theless, due to their low computational complexity, SVMs classification accuracy was achieved with the multiclass remain one of the most popular types of classifiers for two-layer LSTM at 93.27%. EEG-based BCI, in particular for online scenarios. Not- The 4 binary FBCSP-SVM classifiers performed as fol - withstanding that adaptive implementations of SVM have lows: 80.45% for ‘Guided Breathing’ vs ‘Stressor’, 87.88% been found to be superior to their static counterparts, for ‘Baseline’ vs ‘Stressor’, 80.57% for ‘Unguided Breath- they often require fully retraining the classifier with new ing’ vs ‘Guided Breathing’, and 78.28% for ‘Unguided incoming data, resulting in a much higher computa- Breathing’ vs ‘Stressor’. The FBCSP-SVM performed best tional complexity and thus a lack of online applicability when classifying between the pre-task onset resting state [53]. Conversely, with deep learning methods adaptation epoch (‘Baseline’) and the stress induction (‘Stressor’) can be achieved by retraining the input layer with new Sundar esan et al. Brain Inf. (2021) 8:13 Page 9 of 12 incoming data. LSTMs in particular are inherently adap- as a proxy for anxiety. Second, our experiment was tive and thus well suited for real-time scenarios, as their designed to induce anxiety as efficiently as possible and predictions are conditioned by past input. In addition, thus minimize the time under stress to avoid any undue unlike SVMs, deep learning networks can automatically strain on the participants. Conversely, more time was adjust and optimize their parameters, essentially alleviat- required for relaxation to set in and the breathing rate to ing the need for feature extraction and requiring far less normalize; as a result the time for the mental arithmetic processing and prior knowledge regarding the original task and the guided or unguided breathing differed. Thus, EEG dataset [55, 65]. Lastly, while some multiclass SVMs learning models were trained on an unbalanced dataset, have been found to outperform neural networks [89], our with more ‘Unguided Breathing’ and ‘Guided Breath- attempts with multiclass FBCSP-SVMs produced incon- ing’ EEG samples than ‘Stressor’ and ‘Baseline’ samples, sistent results with accuracies ranging between chance- with the potential of artificially inflating model accuracy. level and 90%. However, this is unlikely to be a concern for the two-layer With regard to the multiclass deep learning models, the LSTM RNN model, which exhibited high sensitivity and Deep ConvNet CNN performed with an overall accuracy specificity metrics across all classes. Lastly, it should be of 58.80%, the Shallow ConvNet CNN with 62.84%, the noted that a potential drawback of LSTM RNNs, and EEGNet CNN with 61.18%, the LSTM-FCN with 62.97%, of deep learning algorithms in general, is over-reliance the one-layer LSTM with 73.53%, the two-layer LSTM upon large datasets. To this regard, the same classifica - with 93.27% and the three-layer LSTM with 72.26%. The tions performed with smaller datasets including only high classification accuracies achieved with the LSTM 2 or 3 conditions led to poorer performance (data not architecture presumably is a result of its ability to learn shown). However, the experimental 2-layer LSTM accu- time dependencies within the data. Indeed, the retention racy metrics were likely not impacted by a smaller sample property is useful in mental state monitoring, as con- size and were indicative of the model’s real-world per- sidering the past activations of the EEG can drastically formance due to every trained model’s very high dem- improve the prediction of target variables and the brain onstrated sensitivity, specificity, and predictive ability; activity patterns leading up to, and associated with, spe- we found the 2-layer LSTM models were not only suc- cific cognitive states. The inherent nature of deep learn - cessful in identifying true positives across all classes but ing models, with hidden layers obscuring intermediate also when rejecting false positives in a statistically signifi - processes occurring within the models, makes it chal- cant manner in each and every one of our test subjects. lenging to definitively identify the exact cause for the Indeed, one major drawback of deep learning is the need reduction in performance with the addition of a third for large amounts of data, an issue we aim to remedy in a LSTM layer. However, it is generally understood that future study involving a much larger set of participants, stacking LSTM layers can render the model prone to both neurotypical and ASD, as well as a more diverse set overfitting [90] as well as the vanishing gradient prob - of stress induction tasks. Given that we have identified lem, in which network weights fail to update significantly a viable classifier for the monitoring of cognitive states over time and model training becomes stagnant [91, 92], related to anxiety, the goal of forthcoming studies will phenomena that could explain the lower accuracy of the be to refine and validate the two-layer LSTM RNN deep three-layer LSTM. Empirically, it has been shown that a learning model for prospective implementation in a per- second LSTM layer often provides a significant boost in sonalized pBCI. Such a system will be capable of moni- classification accuracy over a single LSTM [84, 93, 94]; toring for periods of stress and hone in on an individual’s however, the addition of a third LSTM layer or more can optimal respiration patterns by adapting the breathing have little to no effect on performance [55, 93, 94], and entrainment parameters in a closed-loop manner. in some cases additional layers can hinder model training In summary, the goal of this study was to compare and convergence, and in turn degrade performance [91]. several learning models or classifiers on their abil - Consequently, our leading hypothesis finds that the mar - ity to assess mental stress levels from EEG recordings ginal increase in model complexity between the two-layer performed on adolescent students to determine the and three-layer LSTM further complicated model train- feasibility of an EEG-based BCI capable of real-time ing while adding comparatively limited improvements, identification and the mitigation of anxiety through leading to a net loss in model performance. optimized respiration entrainment. Of the different There are some caveats to consider in the interpreta - classifiers we compared, two-layer LSTM yielded the tion of our results. First, given that anxiety varies signifi - highest classification accuracy (93.27%), opening new cantly with context and individual, and cannot therefore avenues of decoding covert mental states for BCI- be induced reliably and equivalently across participants, based neuroadaptive applications to benefit youth with we utilized mental stress induction via mental arithmetic autism. Sundaresan et al. Brain Inf. (2021) 8:13 Page 10 of 12 Acknowledgements 3. Maddox BB, White SW (2015) Comorbid social anxiety disorder in adults We would like to thank The Nueva School, Learning Farm Educational with autism spectrum disorder. J Autism Dev Disord 45:3949–3960. Resources, OpenBCI, and Muvik Labs for their kind and invaluable support. https:// doi. org/ 10. 1007/ s10803‑ 015‑ 2531‑5 The authors would also like to thank the Naturalia & Biologia Foundation for 4. van Steensel FJA, Heeman EJ (2017) Anxiety levels in children with autism financial assistance for traveling and attending meetings. spectrum disorder: a meta‑analysis. J Child Fam Stud 26:1753–1767. https:// doi. org/ 10. 1007/ s10826‑ 017‑ 0687‑7 Authors’ contributions 5. Duvekot J, van der Ende J, Verhulst FC, Greaves‑Lord K (2018) Examining AS, BP, and SC contributed equally to this work. AM, AS, AV‑ C, and VG con‑ bidirectional effects between the autism spectrum disorder (ASD) core ceived of the study idea. AM supervised the project with the help of AV‑ C. AM symptom domains and anxiety in children with ASD. J Child Psychol and VG designed and implemented the experimental paradigm. AS, BP, SC, Psychiatry 59(3):277–284. https:// doi. org/ 10. 1111/ jcpp. 12829 and AM carried out the experiment, developed the theory, and performed 6. Tomczak MT, Wójcikowski M, Pankiewicz B, Łubiński J, Majchrowicz J, the computations and analyses. SC processed the experimental data and Majchrowicz D, Walasiewicz A, Kiliński T, Szczerska M (2020) Stress moni‑ implemented the machine learning algorithms with the close help of AS and toring system for individuals with autism spectrum disorders. IEEE Access the guidance of AM. AS took the lead in writing the manuscript under the 8:228236–228244. https:// doi. org/ 10. 1109/ ACCESS. 2020. 30456 33 supervision of AM and support from BP, SC, and AV‑ C. All authors provided 7. Preece D, Howley M (2018) An approach to supporting young people critical feedback and helped shape the research, analysis and manuscript. All with autism spectrum disorder and high anxiety to re‑ engage with authors read and approved the final manuscript. formal education ‑ the impact on young people and their families. Int J Adolesc Youth 23:468–481. https:// doi. org/ 10. 1080/ 02673 843. 2018. 14336 Funding 95 The contributions of Dr. Adrien Martel and Prof. Valero‑ Cabré have been 8. Kerns CM, Kendall PC, Zickgraf H et al (2015) Not to be overshadowed or supported by research grants ANR (French National Research Agency) projet overlooked: functional impairments associated with comorbid anxiety Générique ‘OSCILOSCOPUS’ and Flag‑ era‑ JTC‑2017 CAUSALTOMICS to AV ‑ C. disorders in youth with ASD. Behav Ther 46:29–39. https:// doi. org/ 10. Dr. Martel is currently supported by a Marie Curie Fellowship (EU project 1016/j. beth. 2014. 03. 005 898813‑ CLONESA‑DLV ‑898813). 9. Clark BG, Magill‑Evans JE, Koning CJ (2015) Youth with autism spectrum disorders: self‑ and proxy‑reported quality of life and adaptive function‑ Availability of data and materials ing. Focus Autism Other Dev Disabl 30:57–64. https:// doi. org/ 10. 1177/ The informed consent forms signed by the study participants and their legal 10883 57614 522289 guardians assured anonymity and utmost confidentiality of participant data, 10. Ikeda E, Hinckson E, Krägeloh C (2014) Assessment of quality of life in so it is not possible to make the data publicly available at this time. children and youth with autism spectrum disorder: a critical review. Qual Life Res 23:1069–1085. https:// doi. org/ 10. 1007/ s11136‑ 013‑ 0591‑6 11. Cadman T, Eklund H, Howley D et al (2012) Caregiver burden as people Declarations with autism spectrum disorder and attention‑ deficit/hyperactivity disor ‑ der transition into adolescence and adulthood in the United Kingdom. J Ethics approval and consent to participate Am Acad Child Adolesc Psychiatry 51:879–888. https:// doi. org/ 10. 1016/j. The study was approved by a panel composed of an educator at Learning jaac. 2012. 06. 017 Farm Educational Resources (Menlo Park, CA), and an administrator and a 12. Smith IC, Ollendick TH, White SW (2019) Anxiety moderates the influence licensed child mental health professional at The Nueva School (San Mateo, of ASD severity on quality of life in adults with ASD. Res Autism Spectr CA). All participants and their parents or legal guardians were informed exten‑ Disord 62:39–47. https:// doi. org/ 10. 1016/j. rasd. 2019. 03. 001 sively about the study objectives and procedures, and all provided written 13. van Steensel FJA, Bögels SM, Dirksen CD (2012) Anxiety and quality of life: informed consent. clinically anxious children with and without autism spectrum disorders compared. J Clin Child Adolesc Psychol 41(6):731–738. https:// doi. org/ 10. Consent for publication 1080/ 15374 416. 2012. 698725 All authors have read the final manuscript and consented for its publication. 14. Vasa RA, Keefer A, Reaven J et al (2018) Priorities for advancing research on youth with autism spectrum disorder and co‑ occurring Competing interests anxiety. J Autism Dev Disord 48:925–934. https:// doi. org/ 10. 1007/ The authors declare that they have no competing interests. s10803‑ 017‑ 3320‑0 15. Hufnagel C, Chambres P, Bertrand PR, Dutheil F (2017) The need for Author details objective measures of stress in autism. Front Psychol 8:64. https:// doi. org/ 1 2 The Nueva School, San Mateo, CA 94033, USA. Muvik Labs, LLC, Locust 10. 3389/ fpsyg. 2017. 00064 Valley, NY 11560, USA. Center for Computer Research in Music and Acous‑ 16. Simpson K, Adams D, Wheeley E, Keen D (2020) Parent perspectives on tics, Stanford University, Stanford, CA 94305, USA. Department of Anatomy the presentation, triggers, impact, and support of anxiety in young chil‑ and Neurobiology, Laboratory of Cerebral Dynamics, Boston University School dren on the autism spectrum. J Child Fam Stud 29(2):572–582. https:// of Medicine, Boston, MA 02118, USA. Cognitive Neuroscience and Informa‑ doi. org/ 10. 1007/ s10826‑ 019‑ 01576‑5 tion Technology Research Program, Open University of Catalonia (UOC), 17. Adams D, Emerson L‑M (2020) The impact of anxiety in children on Barcelona, Spain. Causal Brain Dynamics, Plasticity and Rehabilitation Team, the autism spectrum. J Autism Dev Disord. https:// doi. org/ 10. 1007/ Frontlab, Brain and Spine Institute, ICM, CNRS UMR, Paris 7225, France. s10803‑ 020‑ 04673‑3 18. Blanck P, Perleth S, Heidenreich T et al (2018) Eec ff ts of mindfulness exer ‑ Received: 19 November 2020 Accepted: 31 May 2021 cises as stand‑alone intervention on symptoms of anxiety and depres‑ sion: Systematic review and meta‑analysis. Behav Res Ther 102:25–35. https:// doi. org/ 10. 1016/j. brat. 2017. 12. 002 19. Jayawardena R, Ranasinghe P, Ranawaka H et al (2020) Exploring the therapeutic benefits of pranayama (yogic breathing): a systematic review. References Int J Yoga 13:99–110. https:// doi. org/ 10. 4103/ ijoy. IJOY_ 37_ 19 1. American Psychiatric Association (2013) Diagnostic and statistical manual 20. Chen Y‑F, Huang X ‑ Y, Chien C‑H, Cheng J‑F (2017) The effectiveness of of mental disorders (DSM‑5). American Psychiatric Association, Virginia diaphragmatic breathing relaxation training for reducing anxiety. Per‑ 2. van Steensel FJA, Bögels SM, Perrin S (2011) Anxiety disorders in children spect Psychiatr Care 53:329–336. https:// doi. org/ 10. 1111/ ppc. 12184 and adolescents with autistic spectrum disorders: a meta‑analysis. 21. Russo MA, Santarelli DM, O’Rourke D (2017) The physiological effects Clin Child Fam Psychol Rev 14:302–317. https:// doi. org/ 10. 1007/ of slow breathing in the healthy human. Breathe (Sheff ) 13:298–309. s10567‑ 011‑ 0097‑0 https:// doi. org/ 10. 1183/ 20734 735. 009817 22. Steffen PR, Austin T, DeBarros A, Brown T (2017) The Impact of Reso ‑ nance Frequency Breathing on Measures of Heart Rate Variability, Blood Sundar esan et al. Brain Inf. (2021) 8:13 Page 11 of 12 Pressure, and Mood. Frontiers in Public Health 5:222. https:// doi. org/ 10. 43. Schultze‑Kraft M, Dähne S, Gugler M, Curio G, Blankertz B (2016) Unsuper ‑ 3389/ fpubh. 2017. 00222 vised classification of operator workload from brain signals. J Neural Eng 23. Hopper SI, Murray SL, Ferrara LR, Singleton JK (2019) Eec ff tiveness of dia‑ 13(3):036008. https:// doi. org/ 10. 1088/ 1741‑ 2560/ 13/3/ 036008 phragmatic breathing for reducing physiological and psychological stress 44. Arico P, Borghini G, Di Flumeri G et al (2016) Adaptive automation trig‑ in adults: a quantitative systematic review. JBI Database Syst Rev Imple‑ gered by EEG‑based mental workload index: a passive brain‑ computer ment Rep 17:1855–1876. https:// doi. org/ 10. 11124/ JBISR IR‑ 2017‑ 003848 interface application in realistic air traffic control environment. Front Hum 24. Xiao Y, Ma F, Lv Y et al (2015) Sustained attention is associated with error Neurosci 10:539. https:// doi. org/ 10. 3389/ fnhum. 2016. 00539 processing impairment: evidence from mental fatigue study in four‑ 45. Martel A, Dähne S, Blankertz B (2014) EEG predictors of covert vigilant choice reaction time task. PLoS One 10:e0117837. https:// doi. org/ 10. attention. J Neural Eng 11:035009. https:// doi. org/ 10. 1088/ 1741‑ 2560/ 1371/ journ al. pone. 01178 3711/3/ 035009 25. Ma X, Yue Z‑ Q, Gong Z‑ Q et al (2017) The effect of diaphragmatic 46. Van Benthem KD, Cebulski S, Herdman CM, Keillor J (2018) An EEG brain‑ breathing on attention, negative affect and stress in healthy adults. Front computer interface approach for classifying vigilance states in humans: a Psychol 8:874. https:// doi. org/ 10. 3389/ fpsyg. 2017. 00874 gamma band focus supports low misclassification rates. Int J Hum‑ Com‑ 26. Kim H‑ G, Cheon E‑ J, Bai D‑S et al (2018) Stress and heart rate variability: a put Int 34:226–237. https:// doi. org/ 10. 1080/ 10447 318. 2017. 13429 42 meta‑analysis and review of the Literature. Psychiatry Investig 15:235– 47. Liu Y, Tang Y, Yao D, Zhou X (2019) A vigilance detection method for HSR 245. https:// doi. org/ 10. 30773/ pi. 2017. 08. 17 drivers based on convolutional neural networks and wireless wearable 27. Cho H, Ryu S, Noh J, Lee J (2016) The effectiveness of daily mindful EEG collection technology. In: Transportation Research Board 98th annual breathing practices on test anxiety of students. PLoS One 11:e0164822. meeting, Washington DC, United States, 13–17 January 2019 https:// doi. org/ 10. 1371/ journ al. pone. 01648 22 48. Zhang X, Li J, Liu Y et al (2017) Design of a fatigue detection system for 28. Zaccaro A, Piarulli A, Laurino M et al (2018) How breath‑ control can high‑speed trains based on driver vigilance using a wireless wearable change your life: a systematic review on psycho‑physiological correlates EEG. Sensors. https:// doi. org/ 10. 3390/ s1703 0486 of slow breathing. Front Hum Neurosci 12:353. https:// doi. org/ 10. 3389/ 49. Roy RN, Bonnet S, Charbonnier S, Campagne A (2013) Mental fatigue fnhum. 2018. 00353 and working memory load estimation: interaction and implications for 29. Freeman FG, Mikulka PJ, Prinzel LJ, Scerbo MW (1999) Evaluation of an EEG‑based passive BCI. Conf Proc IEEE Eng Med Biol Soc 2013:6607–6610. adaptive automation system using three EEG indices with a visual track‑https:// doi. org/ 10. 1109/ EMBC. 2013. 66110 70 ing task. Biol Psychol 50:61–76. https:// doi. org/ 10. 1016/ s0301‑ 0511(99) 50. Sitaram R, Ros T, Stoeckel L et al (2017) Closed‑loop brain training: the 00002‑2 science of neurofeedback. Nat Rev Neurosci 18:86–100. https:// doi. org/ 30. Zander TO, Kothe C (2011) Towards passive brain‑ computer interfaces: 10. 1038/ nrn. 2016. 164 applying brain‑ computer interface technology to human‑machine 51. White EK, Groeneveld KM, Tittle RK et al (2017) Combined neurofeedback systems in general. J Neural Eng 8:025005. https:// doi. org/ 10. 1088/ 1741‑ and heart rate variability training for individuals with symptoms of anxi‑ 2560/8/ 2/ 025005 ety and depression: a retrospective study. NR 4:37–55. https:// doi. org/ 10. 31. Krol LR, Zander TO (2017) Passive BCI‑based neuroadaptive systems. 15540/ nr.4. 1. 37 GBCIC, Austria 52. Crivelli D, Fronda G, Balconi M (2019) Neurocognitive enhancement 32. Berka C, Levendowski DJ, Cvetinovic MM, Petrovic MM, Davis G, Lumicao effects of combined mindfulness‑neurofeedback training in sport. Neuro ‑ MN, Zivkovic VT, Popovic MV, Olmstead R (2004) Real‑time analysis of eeg science 412:83–93. https:// doi. org/ 10. 1016/j. neuro scien ce. 2019. 05. 066 indexes of alertness, cognition, and memory acquired with a wireless EEG 53. Lotte F, Bougrain L, Cichocki A et al (2018) A review of classification headset. Int J Hum‑ comput Int 17(2):151–170. https:// doi. org/ 10. 1207/ algorithms for EEG‑based brain‑ computer interfaces: a 10 year update. J s1532 7590i jhc17 02_3 Neural Eng 15:031005. https:// doi. org/ 10. 1088/ 1741‑ 2552/ aab2f2 33. McFarland DJ, Wolpaw JR (2017) EEG‑based brain‑ computer interfaces. 54. Roy Y, Banville H, Albuquerque I et al (2019) Deep learning‑based electro ‑ Curr Opin Biomed Eng 4:194–200. https:// doi. org/ 10. 1016/j. cobme. 2017. encephalography analysis: a systematic review. J Neural Eng 16:051001. 11. 004https:// doi. org/ 10. 1088/ 1741‑ 2552/ ab260c 34. Wolpaw JR, Birbaumer N, McFarland DJ et al (2002) Brain‑ computer inter‑ 55. Craik A, He Y, Contreras‑ Vidal JL (2019) Deep learning for electroen‑ faces for communication and control. Clin Neurophysiol 113:767–791. cephalogram (EEG) classification tasks: a review. J Neural Eng 16:031001. https:// doi. org/ 10. 1016/ S1388‑ 2457(02) 00057‑3https:// doi. org/ 10. 1088/ 1741‑ 2552/ ab0ab5 35. Shih JJ, Krusienski DJ, Wolpaw JR (2012) Brain‑ computer interfaces in 56. Yang H, Han J, Min K (2019) A multi‑ column CNN model for emotion medicine. Mayo Clin Proc 87:268–279. https:// doi. org/ 10. 1016/j. mayocp. recognition from EEG signals. Sensors. https:// doi. org/ 10. 3390/ s1921 4736 2011. 12. 008 57. Salama ES, El‑Khoribi RA, Shoman ME, Wahby MA, (2018) EEG‑based 36. Brunner C, Birbaumer N, Blankertz B et al (2015) BNCI Horizon 2020: emotion recognition using 3D convolutional neural networks. IJACSA. towards a roadmap for the BCI community. BCI 2:1–10. https:// doi. org/ 10. https:// doi. org/ 10. 14569/ IJACSA. 2018. 090843 1080/ 23262 63X. 2015. 10089 56 58. Wang Y, McCane B, McNaughton N et al (2019) AnxietyDecoder: an 37. Müller K‑R, Tangermann M, Dornhege G et al (2008) Machine learning EEG‑based anxiety predictor using a 3‑D convolutional neural network. for real‑time single ‑trial EEG‑analysis: from brain‑ computer interfacing to In: 2019 international joint conference on neural networks (IJCNN), Buda‑ mental state monitoring. J Neurosci Methods 167:82–90. https:// doi. org/ pest, Hungary, 14–19 July 2019, pp 1–8 10. 1016/j. jneum eth. 2007. 09. 022 59. Keng Ang K, Yang Chin Z, Zhang H, Guan C (2008) Filter bank com‑ 38. Arico P, Borghini G, Di Flumeri G et al (2017) Passive BCI in operational mon spatial pattern (FBCSP) in brain‑ computer interface. In: 2008 IEEE environments: insights, recent advances, and future trends. IEEE Trans international joint conference on neural networks (IEEE world congress Biomed Eng 64:1431–1436. https:// doi. org/ 10. 1109/ TBME. 2017. 26948 56 on computational intelligence), Hong Kong, China, 1–8 June 2008, pp 39. Clerc M, Bougrain L, Lotte F (2016) Brain‑ computer interfaces 1. Wiley‑ 2390–2397 ISTE, London 60. Schirrmeister RT, Springenberg JT, Fiederer LDJ et al (2017) Deep learning 40. Lotte F, Congedo M, Lécuyer A et al (2007) A review of classification algo ‑ with convolutional neural networks for EEG decoding and visualization. rithms for EEG‑based brain‑ computer interfaces. J Neural Eng 4:R1–R13. Hum Brain Mapp 38:5391–5420. https:// doi. org/ 10. 1002/ hbm. 23730 https:// doi. org/ 10. 1088/ 1741‑ 2560/4/ 2/ R01 61. Al‑Shargie F, Kiguchi M, Badruddin N, Dass SC, Hani AFM, Tang TB (2016) 41. Blankertz B, Curio G, Müller K‑R (2002) Classifying single trial EEG: towards Mental stress assessment using simultaneous measurement of EEG and brain computer interfacing. In: Dietterich TG, Becker S, Ghahramani Z fNIRS. Biomed Opt Express 7(10):3882–3898. https:// doi. org/ 10. 1364/ (eds) Advances in neural information processing systems 14. MIT Press, BOE.7. 003882 Cambridge, pp 157–164 62. Gaikwad P, Paithane AN (2017) Novel approach for stress recognition 42. Appriou A, Cichocki A, Lotte F (2018) Towards robust neuroadaptive HCI: using EEG signal by SVM classifier. In: 2017 international conference on exploring modern machine learning methods to estimate mental work‑ computing methodologies and communication (ICCMC), Erode, India, load from EEG signals. In: Extended abstracts of the 2018 CHI conference 18–19 July 2017, pp 967–971 on human factors in computing systems. Paper No.: LBW615. ACM, New York, pp 1–6. https:// doi. org/ 10. 1145/ 31704 27. 31886 17 Sundaresan et al. Brain Inf. (2021) 8:13 Page 12 of 12 63. Al‑Shargie F, Tang TB, Badruddin N, Kiguchi M (2018) Towards multilevel 80. Gramfort A, Luessi M, Larson E et al (2014) MNE software for processing mental stress assessment using SVM with ECOC: an EEG approach. Med MEG and EEG data. Neuroimage 86:446–460. https:// doi. org/ 10. 1016/j. Biol Eng Comput 56:125–136. https:// doi. org/ 10. 1007/ s11517‑ 017‑ 1733‑8neuro image. 2013. 10. 027 64. Saeed SMU, Anwar SM, Khalid H et al (2020) EEG based classification of 81. Vanhatalo S, Voipio J, Kaila K (2005) Full‑band EEG (FbEEG): an emerging long‑term stress using psychological labeling. Sensors. https:// doi. org/ 10. standard in electroencephalography. Clin Neurophysiol 116:1–8. https:// 3390/ s2007 1886doi. org/ 10. 1016/j. clinph. 2004. 09. 015 65. Schmidhuber J (2015) Deep learning in neural networks: an overview. 82. Duan K‑B, Keerthi SS (2005) Which is the best multiclass SVM method? An Neural Netw 61:85–117. https:// doi. org/ 10. 1016/j. neunet. 2014. 09. 003 empirical study. Multiple classifier systems. Springer, Berlin Heidelberg, pp 66. Lawhern VJ, Solon AJ, Waytowich NR et al (2018) EEGNet: a compact 278–285 convolutional neural network for EEG‑based brain‑ computer interfaces. J 83. Ramoser H, Müller‑ Gerking J, Pfurtscheller G (2000) Optimal spatial Neural Eng 15:056013. https:// doi. org/ 10. 1088/ 1741‑ 2552/ aace8c filtering of single trial EEG during imagined hand movement. IEEE Trans 67. Hochreiter S, Schmidhuber J (1997) Long short‑term memory. Neural Rehabil Eng 8:441–446. https:// doi. org/ 10. 1109/ 86. 895946 Comput 9:1735–1780. https:// doi. org/ 10. 1162/ neco. 1997.9. 8. 1735 84. Tsiouris KM, Pezoulas VC, Zervakis M, Konitsiotis S, Koutsouris DD, Fotiadis 68. Xing X, Li Z, Xu T et al (2019) SAE+LSTM: a new framework for emotion DI (2018) A long short‑term memory deep learning network for the recognition from multi‑ channel EEG. Front Neurorobot 13:37. https:// doi. prediction of epileptic seizures using EEG signals. Comput Biol Med org/ 10. 3389/ fnbot. 2019. 00037 99:24–37. https:// doi. org/ 10. 1016/j. compb iomed. 2018. 05. 019 69. Alhagry S, Fahmy AA, El‑Khoribi RA (2017) Emotion Recognition based 85. Srivastava N, Hinton G, Krizhevsky A et al (2014) Dropout: a simple way to on EEG using LSTM recurrent neural network. IJACSA. https:// doi. org/ 10. prevent neural networks from overfitting. J Mach Learn Res 15:1929–1958 14569/ IJACSA. 2017. 081046 86. Kingma DP, Ba J (2014) Adam: a method for stochastic optimization. In: 70. Yang Y, Wu Q, Qiu M et al (2018) Emotion recognition from multi‑ channel 3rd international conference for learning representations, San Diego, EEG through parallel convolutional recurrent neural network. In: 2018 2015. arXiv: 1412. 6980 international joint conference on neural networks (IJCNN), Rio de Janeiro, 87. Faust O, Hagiwara Y, Hong TJ et al (2018) Deep learning for healthcare Brazil, 8–13 July 2018, pp 1–7 applications based on physiological signals: A review. Comput Methods 71. Li Y, Huang J, Zhou H, Zhong N (2017) Human emotion recognition with Programs Biomed 161:1–13. https:// doi. org/ 10. 1016/j. cmpb. 2018. 04. 005 electroencephalographic multidimensional features by hybrid deep 88. Vahid A, Mückschel M, Stober S et al (2020) Applying deep learning to neural networks. NATO Adv Sci Inst Ser E Appl Sci 7:1060. https:// doi. org/ single‑trial EEG data provides evidence for complementary theo ‑ 10. 3390/ app71 01060 ries on action control. Commun Biol 3:112. https:// doi. org/ 10. 1038/ 72. Karim F, Majumdar S, Darabi H, Chen S (2018) LSTM fully convolutional s42003‑ 020‑ 0846‑z networks for time series classification. IEEE Access 6:1662–1669. https:// 89. Güler I, Ubeyli ED (2007) Multiclass support vector machines for EEG‑ doi. org/ 10. 1109/ ACCESS. 2017. 27799 39 signals classification. IEEE Trans Inf Technol Biomed 11:117–126. https:// 73. Borthakur D, Grace V, Batchelor P et al (2019) Fuzzy C‑means cluster ‑doi. org/ 10. 1109/ titb. 2006. 879600 ing and sonification of HRV features. In: 2019 IEEE/ACM international 90. Dai Yin, Jha, (2020) Grow and prune compact, fast, and accurate LSTMs. conference on connected health: applications, systems and engineering IEEE Trans Comput 69:441–452. https:// doi. org/ 10. 1109/ TC. 2019. 29544 95 technologies (CHASE), Arlington, VA, USA, 25–27 September 2019, pp 91. Zhang Y, Chen G, Yu D, Yao K, Khudanpur S, Glass J (2016) Highway long 53–57 short‑term memory RNNS for distant speech recognition. In: 2016 IEEE 74. Spielberger CD (1973) Manual for the state‑trait inventory for children. international conference on acoustics, speech and signal processing Palo Alto, Santa Clara (ICASSP), Shanghai, China, 20–25 March 2016, pp 5755–5759. https:// doi. 75. Simon DM, Corbett BA (2013) Examining associations between anxiety org/ 10. 1109/ ICASSP. 2016. 74727 80 and cortisol in high functioning male children with autism. J Neurodev 92. Hsu W‑N, Zhang Y, Glass J (2016) A prioritized grid long short ‑term Disord 5:32. https:// doi. org/ 10. 1186/ 1866‑ 1955‑5‑ 32 memory RNN for speech recognition. SLT 2016:467–473. https:// doi. org/ 76. Dedovic K, Renwick R, Mahani NK et al (2005) The montreal imaging 10. 1109/ SLT. 2016. 78463 05 stress task: using functional imaging to investigate the effects of perceiv‑ 93. Tseng P‑H, Urpi NA, Lebedev M, Nicolelis M (2019) Decoding movements ing and processing psychosocial stress in the human brain. J Psychiatry from cortical ensemble activity using a long short‑term memory recur ‑ Neurosci 30:319–325 rent network. Neural Comput 31(6):1085–1113. https:// doi. org/ 10. 1162/ 77. Shilton AL, Laycock R, Crewther SG (2017) The maastricht acute stress test neco_a_ 01189 (MAST ): physiological and subjective responses in anticipation, and post‑ 94. Wang P, Jiang A, Liu X, Shang J, Zhang L (2018) LSTM‑based EEG clas‑ stress. Front Psychol 8:567. https:// doi. org/ 10. 3389/ fpsyg. 2017. 00567 sification in motor imagery tasks. IEEE Trans Neural Syst Rehabil Eng 78. Szulczewski MT (2019) Training of paced breathing at 0.1 Hz improves 26(11):2086–2095. https:// doi. org/ 10. 1109/ TNSRE. 2018. 28761 29 CO2 homeostasis and relaxation during a paced breathing task. PLoS One 14. https:// doi. org/ 10. 1371/ journ al. pone. 02185 50 Publisher’s Note 79. Lin IM, Tai LY, Fan SY (2014) Breathing at a rate of 5.5 breaths per minute Springer Nature remains neutral with regard to jurisdictional claims in pub‑ with equal inhalation‑to ‑ exhalation ratio increases heart rate variability. lished maps and institutional affiliations. Int J Psychophysiol 91:206–211. https:// doi. org/ 10. 1016/j. ijpsy cho. 2013. 12. 006

Journal

Brain Informatics – Springer Journals

Published: Jul 13, 2021

Keywords: Mental stress; Autism; EEG; Deep learning; Breathing entrainment

There are no references for this article.