Combining EEG signal processing with supervised methods for Alzheimer’s patients classification

Combining EEG signal processing with supervised methods for Alzheimer’s patients classification Background: Alzheimer’s Disease (AD) is a neurodegenaritive disorder characterized by a progressive dementia, for which actually no cure is known. An early detection of patients affected by AD can be obtained by analyzing their electroencephalography (EEG) signals, which show a reduction of the complexity, a perturbation of the synchrony, and a slowing down of the rhythms. Methods: In this work, we apply a procedure that exploits feature extraction and classification techniques to EEG signals, whose aim is to distinguish patient affected by AD from the ones affected by Mild Cognitive Impairment (MCI) and healthy control (HC) samples. Specifically, we perform a time-frequency analysis by applying both the Fourier and Wavelet Transforms on 109 samples belonging to AD, MCI, and HC classes. The classification procedure is designed with the following steps: (i) preprocessing of EEG signals; (ii) feature extraction by means of the Discrete Fourier and Wavelet Transforms; and (iii) classification with tree-based supervised methods. Results: By applying our procedure, we are able to extract reliable human-interpretable classification models that allow to automatically assign the patients into their belonging class. In particular, by exploiting a Wavelet feature extraction we achieve 83%, 92%, and 79% of accuracy when dealing with HC vs AD, HC vs MCI, and MCI vs AD classification problems, respectively. Conclusions: Finally, by comparing the classification performances with both feature extraction methods, we find out that Wavelets analysis outperforms Fourier. Hence, we suggest it in combination with supervised methods for automatic patients classification based on their EEG signals for aiding the medical diagnosis of dementia. Keywords: Alzheimer’s disease, Feature extraction, Electroencephalography signals, Classification Background dementiarepresentsoneofthemajorplaguefor themod- Dementia is a broad group of brain disorders leading to ern society. The most widespread cause of dementia is the a cognitive impairment because of a gradual dysfunction Alzheimer’s disease (AD), which involves serious mem- and death of brain cells. The World Alzheimer Report ory loss, cognitive impairment, and behavioural changes. 2015 has been estimated that 36 million people were liv- Thus, AD interferes with daily, social and professional ing with dementia in 2010, nearly doubling every 20 years functioning of patients, also affecting the daily life of their to 66 million by 2030 and to 115 million by 2050 [1]. families [2]. The intermediate stage between the normal Given the continuous growth of incidence of this illness, cognitive deficit due to aging and dementia is defined as Mild Cognitive Impairment (MCI). Several symptoms dis- tinguish MCI, but the loss of memory is a risk factor to *Correspondence: giulia.fiscon@iasi.cnr.it Giulia Fiscon and Emanuel Weitschek contributed equally to this work. develop AD [3]. In Europe, only 50% of the patients with Institute of Systems Analysis and Computer Science A. Ruberti (IASI), National dementia receive a diagnosis by a specialist centre, and Research Council (CNR), Via dei Taurini 19, 00185 Rome, Italy tests for dementia are carried out after the patient has SysBio Centre for Systems Biology, Rome, Italy Full list of author information is available at the end of the article already started showing symptoms and the disease has © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 2 of 10 progressed [4]. Usually, the process for obtaining a clinical subjects affected by neurodegenerative diseases (e.g., AD) diagnosis for dementia of a patient is mainly based on the or other pathologies (e.g., epilepsy). delivery of a questionnaire in order to assess its cognitive Nevertheless, AD and MCI subjects are characterized abilities. However, a timely diagnosis would facilitate care, by a huge variability and thus discriminating artifacts and reduce the progression of the disease, and improve the patterns similarities to physiological brain activity still patient’s management to alleviate the burden. This might remain a crucial issue. In this regard, EEG signal pro- be achieved through a combination of diagnosis criteria cessing integrated with computational algorithms based and reliable biomarkers. on machine learning methods may contribute to a deeper In the past years, significant progresses have been comprehension of the disease and simplify the work of made to detect the early stages of dementia through bio- neurologists providing an additional tool to diagnose the chemical, genetic, neuroimaging, and neurophysiological stage of dementia [20, 30–33]. biomarkers such as Electroencephalography (EEG) [5– In this paper, we propose a procedure based on EEG- 9]. EEG provides the electrical activity of the brain by signal preprocessing and automatic classification with tracking the connectivity of neurons in the recording supervised learning methods, and its application to dis- sites of the scalp [10], processing it with milliseconds criminate subjects belonging to AD, or MCI, or HC precision. The condition of the brain physiology can be classes. This is an extension of a preliminary work [34]in inferred from the EEG signals recorded, and thus abnor- which we processed an EEG data set composed of 49 AD, malities can be identified through the detection of unusual 37 MCI and 14 healthy controls subjects (HC) by means frequency patterns [11]. Indeed, different rhythms with of a spectrum analysis based on the Fourier Transforma- diverse frequency bands describe the activity of the brain tion, and we automatically classified them with supervised and can be recorded by EEG. Among them, the main ones machine learning methods. Here, we have increased the are alpha (8-13 Hz, 30-50 μV amplitude), beta (13-30 Hz, number of HC subjects of the data set to 23 in order 5-30 μV amplitude), gamma (≥ 30 Hz), delta (0.5-4 Hz), to balance the number of samples for each category. We and theta (4-7 Hz, ≥20μV amplitude). have also improved the EEG-signal preprocessing and Although it is characterized by a lower spatial resolution spectrum analysis techniques through the application of than other neuroimaging techniques, EEG provides high the Wavelet Transform as an efficient method for noise temporal resolution [12]. Moreover, EEG is non-invasive, reduction and feature extraction, obtaining a more reliable ease and faster to use and able to differentiate severity of method to distinguish healthy from diseased subjects. dementia at a lower cost than other imaging devices [13, 14]. Thanks to its reduced costs EEG can be easily imple- Methods mented for population screening to detect pre-clinical We apply a procedure that encompasses the following biomarkers. steps summarized in Fig. 1: (i) data collection (subjects EEG signal analysis may provide useful indications of recruitment, EEG recording) and preprocessing, (ii) fea- the patterns of brain activity and predict the stages of ture extraction (Fast Fourier and Wavelet Analysis), and dementia [15, 16] because of its significant capacity to (iii) classification (supervised machine learning to distin- detect brain rhythm abnormalities, generally correlated guishtheAD,MCI,and HCclasses). with the severity of cognitive impairment [17]. In par- ticular, different clinical studies confirm EEG as suitable Data collection and preprocessing technique to early detect AD [18–20], due to the follow- Subject Recruitment The IRCCS Centro Neurolesi ing effects on EEG signals: reduction of the complexity, “Bonino-Pulejo” enrolled in 2012 and 2013 a total of 109 perturbation of the synchrony, and slowingdown of the subjects: 86 patients affected by dementia (AD, MCI) rhythms [19, 21, 22]. The slowing of the rhythms in the of which 37 men and 49 women, and 23 healthy con- EEG signals of subjects affected by AD can be explained trols samples (HC) of which 13 men and 10 women. The by a gain of the activity in the theta and delta frequency patients have been classified either in AD or MCI, tak- ranges, and a reduction of the activity in the alpha and ing into account the World Health Organization standard. beta frequency ranges [23–26]. The reduction of complex- Subjects capable of undergoing an electroencephalogram ity in the EEG temporal patterns can be explained by a and with a negative anamnesis for neurological comorbid modification of the neural network architecture observed disease have been included. Conversely, subjects under in subjects affected by AD [27, 28] due to loss of neu- pharmacological treatment that could change the activity rons and functional interaction alteration which make the of the brain have been excluded from the study. Patients ac tivity of the brain more predictable, more regular, and are mean aged 78.4 ± 6.4 and 74.1 ± 9.4 years, respectively simplerthaninhealthy controlsamples (HC) [29]. There- for AD and MCI, whereas the mean age of healthy con- fore, we can state that EEG signals related to healthy trols is 65.6 ± 7.9 years. Association between gender and controls subjects can be distinguished from those ones of etiological class (AD, MCI, HC) is not detected by the chi- Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 3 of 10 Fig. 1 Flowchart of the EEG signal analysis procedure square test (p-value > 0.05). Additionally, the difference preprocessed EEG signals, we apply the Fourier and the in terms of age between men and women is not statis- Wavelet Transform for estimating their spectrum [37]. tically significant according to the two-tailed Student’s T Firstly, the Fast Fourier Transform (FFT) is applied to Test (p-value > 0.05 for each class). Thus, the hypothe- each EEG signal of 180 seconds providing M Fourier Coef- sis of homogeneity for age and gender among etiological ficients for each electrode (M equal to 16). Hence, for each classes cannot be rejected. Table 1 provides an overview of sample we obtain 304 features (16 coefficients · 19 elec- the enrolled subjects that can be divided in three main eti- trodes) and we arrange them in a matrix with 109 rows ological classes: (i) patients with Alzheimer’s disease (AD), (referring to the samples) and 305 columns (304 referring (ii) patients with Mild Cognitive Impairment (MCI), and to the features, and one referring to the sample type). (iii) healthy control samples (CT). Secondly, the Discrete Wavelet Transform (DWT) is applied to each EEG signal of 180 seconds providing M Wavelet Coefficients for each electrode (M equal to 48). EEG recording We acquired multi-channel EEG sig- Hence, for each sample we obtain 912 features (48 coef- nals by using 19 electrodes, by setting their placement ficients · 19 electrodes) and we arrange them in a matrix according to the International 10-20 System [35], and by with 109 rows (referring to the samples) and 913 columns exploiting monopolar connections with earlobe electrode (912 referring to the features, and one referring to the landmark [10]. The brain activity of the subjects in rest- sample type). ing condition and closed eyes was measured in terms We provide in Table 2 a schematic representation of the of electrical potential (μV ). We recorded the EEG sig- matrices. nals by capturing 300 seconds with 256 or 1024 sampling The spectral analysis (Wavelet and Fourier) of the EEG frequency (Hz). signals has been performed by using the high level com- o˝ puting language provided by MATLAB R2014a [38]. Preprocessing For each signal we select the central 180 seconds (i.e., from 60 to 240 seconds) to avoid initial and Fourier analysis final EEG recording artifacts. Additionally, to normalize We apply the Fast Fourier Transform (FFT) to obtain the the sampling frequency we convert each signal to 256 Hz. spectrum of the EEG signals [37]. The FFT relies on the Discrete Fourier Transform (DFT) computed as follows: Feature extraction S−1 Extracting features from EEG signals in frequency domain X[ k] = x[ s] e [ s](1) hasbeenproventobeeffectiveforanalyzing theelectrical s=0 brain activity with computational models [31, 36]. Thus, in order to obtain a set of informative features from the Table 2 Schema of the matrix obtained after the feature extraction phase Table 1 Overview of the recruited subjects Sample Coefficient ··· Coefficient Sample type (1,1) (E,E·M) Sample Number of samples (%) Average age (std dev.) in years sample a ··· a HC 1 (1,1) (1,E·M) type Male Female Total Male Female Total sample a ··· a MCI 2 (2,1) (2,E·M) AD 20 (41%) 29 (59%) 49 78.6 (4.1) 78.2 (7.6) 78.4 (6.4) ··· ··· ··· ··· ··· MCI 17 (46%) 20 (54%) 37 75.7 (9.7) 72.7 (9.1) 74.1 (9.4) sample a ··· a AD N (N,1) (N,E·M) HC 13 (56%) 10 (44%) 23 68.1 (6.9) 62.3 (8.3) 65.6 (7.9) N = number of samples, M = number of coefficients, M + 1 = number of features, E Total 50 (46%) 59 (54%) 109 74.9 (8.2) 73.6 (9.9) 74.2 (9.1) = number of electrodes, a = element of the matrix Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 4 of 10 with s representing the s-th sample in the time domain; x In this work, we adopt the DWT in order to per- corresponding to the signal time series (s = 0, 1, 2, ··· , S− form the spectral analysis on the previously described 1); X referring to the representation of th frequency dataset (see section Data collection and preprocessing). domain for the time-series signal x; S representing the the The choice of a simple DWT stems from the need of whole number of samples of the signal x; k correspond- obtaining good performances over an arbitrary num- ing to k-th frequency component (k = 0, 1, ... , S − 1); ber of feature elements per channel and from the sam- jks2π − pling frequency of the input signals (256 Hz). We e [ s] = e referring to the k-th basis function. adopt two types of discrete wavelet families: Daubechies e [ s] is calculated simultaneously during the sampling (db) and Symlets (sym). Daubechies are compactly supported phase. Such a formula yields as output one complex num- orthonormal wavelets [45], while Symlets are symmetrical ber X[ k]for each k component. The output of the FFT wavelets proposed by Daubechies as modifications to the analysis are the Fourier Coefficients arranged in a matrix db family [46]. as showninTable 2. Given a single set of signals, each one is processed Wavelet analysis according to a feature extraction procedure composed of A more effective way for decomposing time and frequency two main phases: noise reduction and feature extraction. of the EEG signal, and for processing it is provided by the Firstly, we perform a noise reduction phase, where each Wavelet Transform (WT). WT is a time-frequency repre- EEG signal is decomposed in n levels (i.e., sub-bands) sentation of the signal, which is decomposed in different by applying a DWT (Symlets order 3 wavelet type). For windows of variable size, i.e., sub-bands. Conversely to every sub-band x an upper and lower threshold value is the FFT, the WT is able to catch the transient features of calculated as: the analyzed signal [39], i.e., it enables to keep both the Thr (x) = avg(x) + 1.5 · stdev(x) (3) temporal (spatial duration) and frequency information of up the signal. Indeed, WT allows to represent when tran- Thr (x) = avg(x) − 1.5 · stdev(x) (4) dwn sient events occur in the signal and with what intensity, as well as the time variations of the frequency contents [40]. The values of each sample s are then compared according Given a signal, WT decomposes it in simpler oscillating to the defined thresholds (3)and (4)and if s > Thr or i up functions called wavelets. A family of wavelets (ψ (t)) s < Thr then s is reduced as follows: s ∗ (Thr (x) − i i i up a,b dwn are derived from a unique mother wavelet ψ(t) by scal- Thr (x))/100. This step is performed in order to obtain dwn ing (dilating and contracting) and by shifting it to different an effective artifact reduction and to avoid possible infor- time positions [40, 41]. mation loss. The artifact removal phase operates on two levels of signal decomposition: level 5 and level 8. We 1 t − b choose these decomposition levels, because their ranges ψ (t) = √ ψ (2) a,b take into account the alpha, theta, beta, and delta band- |a| a widths, which are widely adopted for EEG analysis and In Eq. 2 t is the time variable, a ∈ R \ 0isthe scal- have been proven to be effective when dealing with ing parameter, and b ∈ R is the shifting parameter. The Alzheimer’s diseased patients (see “Background”section wavelets are localized in both time and frequency with for more details). The channel signal is then reconstructed respect to the sinusoidal waves of Fourier, which are bet- with the obtained values, which are given as input to the ter localized in frequency, but infinitely extended in time feature extraction phase. [40]. Additionally, the former are limited in band, i.e., they Secondly, for extracting the features, we adopt the are composed of a defined range of frequencies. Daubechies order 4 (db4) wavelet type with a sampling When dealing with digital signals that are frequency frequency of 256 Hz at decomposition level 5, which has band-limited, the continuous form of WT can be dis- been shown to guarantee a precise feature extraction in cretized according to the sampling theorem [42]. The the brainwaves frequencies [47], and we perform a large Discrete Wavelet Transform (DWT) allows to process dig- set of test with different parameters obtaining lower per- ital signals by keeping enough information in reasonable formances. The feature extraction phase extracts the fol- computational time. A relevant feature of the DWT is the lowing statistical features: mean, standard deviation, and combination with high and low pass filters, through which power spectral density of the wavelet coefficients. All the the signals can be processed to filter the high and low three feature types, representing the frequencies distribu- frequencies in order to compress and reduce the noise tion of the EEG signals, are calculated over the n epochs [43], e.g., hidden artifacts and background noise during of a channel-related signal. This phase makes use of the the EEG signals recording. Indeed, the WT is a well- decomposition levels obtained by applying the DWT to established signal representation and feature extraction the values produced during the noise reduction phase. technique for EEG processing [44]. Our method allows to apply an adaptive, threshold-based Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 5 of 10 noise/artifact removal to the main bandwidths (i.e., alpha, by computing standard statistical metrics, as accuracy, theta, beta, delta). We extract 16 features per channel precision, sensitivity, specificity, and F-measure and by when considering the combination of only two band- adopting a leave-one-out cross validation sampling proce- widths, i.e., alpha - theta or beta - delta, and 12 features per dure [48]. It is worth to note that the classification models channel when taking into account all the four brainwaves. can be adopted to classify new subjects whose diagnosis has not been already assessed and that could constitute The output of the DWT analysis are the Wavelet Coef- ficients arranged in a matrix as shown in Table 2. an independent validation set for further verifying the extracted models. Classification Finally, in order to prove the validity of the extracted We perform a supervised learning analysis in order to models we performed random permutations of class automatically classify the samples to their types (HC, membership for each classification problem and each sig- MCI, AD) by processing their associated features [48, 49]. nal processing technique (Fourier and Wavelet). We test if Supervised learning automatically assigns a sample into our procedure is able to extract meaningful classification a class by inferring a classification model from labeled models regardless of the class partition imposed on the data (training set). Our aim is to extract a human readable training set. This would be verified only in the presence of model specific for each type (HC, MCI, AD) of sample a marked overfitting behavior. containing a small subset of features, e.g., ‘if Wavelet > 0.3 and Wavelet < 0.6 then the sample can be classified Results as MCI”). This model can support clinicians to identify In this section, we provide the classification results rely- pivotal features related to the investigated neurodegen- ing on the features extracted with the Fourier and Wavelet erative disease and to diagnose new cases. In particular, Transforms applied on EEG signals of 180 s. Tables 3 and 4 we address the following classification problems: (i) HC present the results of the Decision Tree classifier consid- vs AD; (ii) HC vs MCI; (iii) MCI vs AD; (iv) HC vs ering the Fourier Transform and the Wavelet Transform, CASE (MCI+AD), where the CASE class is composed of respectively. AD joint to MCI samples in order to test the recogni- In particular, Table 3 presents the results of the Decision tion of the diseased patients with respect to the healthy Tree C4.5 classifier concerning the EEG signals with M = ones. Among the plethora of classification methods we 16 extracted Fourier Coefficients. We obtain 72%, 72%, use Decision Trees classifiers (i.e., C4.5 [50]), because they 80%, 75% of accuracy when dealing with HC vs AD, HC allow to handle noisy datasets and over-fitting with an vs MCI, MCI vs AD, HC vs CASE classification problems, ad-hoc parameters tuning. Additionally, Decision Trees respectively. provide the investigator with a compact, clear, and human Table 4 presents the performance of the Decision Tree readable classification model. C4.5 is an algorithm for the C4.5 classifier concerning the EEG signals processed with generation of decision trees used for classification. A deci- the Wavelet Transform. In all classification tasks, the fea- sion tree is a structure similar to a flow chart, where each ture extraction based on the Wavelet Transform achieves node denotes a test on an attribute, each branch repre- high classification performance in all metrics, obtaining sents a result of a test, and every leaf is labeled by a class. 83%, 92%, 79%, and 73% of accuracy when dealing with Indeed a node with outgoing edges is termed test node HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE classi- and the final nodes are the leaves. fication problems, respectively. In particular, the Wavelet In decision trees the classification model permits to pre- spectral analysis outperforms the Fourier analysis when dict the class of a sample based on its features. The dealing with EEG signals classification of HC vs AD, HC algorithm takes as input a set of classified data (train- vs MCI, and HC vs CASE. Conversely, for MCI vs AD both ing set) and the output is composed by leaf nodes, which define the belonging to a class attribute. Indeed, the path Table 3 Classification performance [%] by using M = 16 Fourier from the root to a specific leaf means that all the pred- Coefficients as features and a leave-one-out sampling with 72, 60, icates applied to the features of the sample are verified. 86, 109 folds for HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE, The validity of the three is verified on a set of labeled respectively samples (test set), but whose class is taken into account HC vs AD HC vs MCI MCI vs AD HC vs CASE only for verification of the class assignments. In this work, we use the J48 Java based implementation of C4.5 avail- Accuracy 72.2 71.7 80.2 74.7 able in the Weka package [51]. In addition, we performed Precision 71.1 78.9 80.2 74.0 a large battery of tests with other families of classifiers Sensitivity 72.2 71.7 80.2 74.7 (function-based, rule-based, naive-based, and Bayesian- Specificity 59.0 79.0 78.5 46.3 based), whose performances are not satisfying and hence F-measure 71.4 71.8 80.1 74.7 not reported. The classification performance is evaluated Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 6 of 10 Table 4 Classification performance [%] by using M = 48 Wavelet MCI vs AD, HC vs CASE permutated classification prob- Coefficients as features and a leave-one-out sampling with 72, 60, lems, respectively. Therefore, we obtain an overall average 86, 109 folds for HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE, classification accuracy of 50.6%. respectively We also test other classification methods, such as function- HC vs AD HC vs MCI MCI vs AD HC vs CASE based, rule-based, naive-based, and Bayesian-based (e.g., Accuracy 83.3 91.7 79.1 73.4 RIPPER [54], SVM [55], the MultiLayer Perceptron [56]), whose performances are not satisfying and hence are Precision 83.3 91.8 79.3 74.7 not reported. For instance, we performed a large num- Sensitivity 83.3 91.7 79.1 73.4 ber of tests with SVMs by tuning the parameters and by Specificity 78.0 91.5 79.1 51.5 setting the complexity, the epsilon for round-off error, F-measure 83.3 91.7 79.1 74.0 the random seed, the tolerance to many different com- bination of values, but results were not above 65% of accuracy. Furthermore, we remark that the adopted C4.5 algo- signal processing methods lead to comparable classifica- rithm provides a classification model, a tree built on tion performances. Wavelet/Fourier Coefficients, from which the investiga- Another validation of the proposed procedure is based tor can derive the corresponding set of EEG electrodes. on the variation of the adopted sampling schema and by Here, the classification trees extracted in our performed applying also a feature selection step with the Information analyses mainly involve the electrodes T, O, F, Fp and the Gain (InfoGain) filter as evaluation measure followed by wavebands alpha, theta, and delta. the Ranker search method, which reached good perfor- Figure 2 depicts an example of such a tree and Fig. 3 mance in our previous study [52]. We perform both 10- shows the scatter plot of the features extracted from this fold cross validation sampling and holdout (90% training tree and related to the considered classes (MCI and HC). and 10% test percentage split) combined with Information Gain filter [53] obtaining again satisfying classification Discussion performance. We observe an improvement of the HC Although different neuroimaging techniques (e.g., Magnetic vs CASE and HC vs MCI classification tasks both for Resonance Imaging, Positron Emission Tomography) can Fourier and Wavelet transforms reaching even more than 80 and 90% of accuracy, respectively. Conversely, a perfor- be used for aiding the diagnosis of dementia providing mance decreasing is observed when distinguishing MCI quantitative data about the brain abnormalities [57, 58], EEG is non-invasive, besides being cheaper, simpler vs AD, probably due to the similarity of the two classes. and faster to use than other imaging devices [13, 14]. Classification results are detailed in Table 5. For this reason, automated EEG signal analysis plays For validating our results and the extracted classifica- an important role in detecting dementia in the early tion models we apply the procedure to data with random stages, as well as in classifying disease severity [59–61]. permutations of class labels. This validation test is per- Supervised learning is doubtless one of the most popular formed on 100 different random permutations for each methods to classify brain disorders with EEG [62–66]. classification problem and for each EEG signal process- Several studies compared the performance of classifi- ing technique (Fourier and Wavelet). In particular, when cation algorithms in terms of sensitivity and specificity, using the Fourier (Wavelet) transform we achieve 56.5% both for the early detection of dementia and for aiding (52.5%), 45.0% (50.5%), 49.5% (45.9%), and 54.5% (50.4%) clinical diagnosis. According to [47], our results show of average accuracy considering HC vs AD, HC vs MCI, that the Wavelet spectral analysis outperforms the Fourier analysis in discriminating EEG of health controls to ones of demented patients. Since EEG may exhibit normal Table 5 Classification performance (Accuracy [%]) by using 10-fold cross validation (CV) sampling and holdout (90% training frequency and may appear similar to normal aged con- and 10% test percentage split) for HC vs AD, HC vs MCI, MCI vs trol subjects during the earliest stages of dementia [67], AD, HC vs CASE, taking into account Wavelet (WT) and Fourier the higher accuracy of C4.5 with WT in distinguishing (FT) Coefficients as features between HC and MCI is a notable result. It is prob- Wavelet Fourier ably due to the fact that WT is suitable for nonsta- 10-fold CV Holdout 10-fold CV Holdout tionary signal like EEG that provide linear combination of the sum of wavelet coefficients and mother wavelet HC vs AD 76.4 71.4 80.6 85.7 with frequency and localization information. In this way, HC vs MCI 93.3 83.3 83.3 83.3 WT is able to detect the slowing of the alpha rhythm, MCI vs AD 66.3 88.9 66.7 77.2 which is more commonly found in intermediate stages HC vs CASE 81.7 81.8 84.4 90.9 of AD [67]. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 7 of 10 Fig. 2 C4.5 tree for HC vs MCI of size 7 with 4 leaves. Each path from the root to a leaf represents a classification rule. Each leaf is associated to a class and two numbers. The first number is the total number of instances recognized by the rule, while the second optional number represents how many ones (if any) are misclassified The choice of using three 2-class classification mod- Furthermore, we tested two other sampling schemes els instead of a single 3-class one is motivated by two (i.e., 10-fold cross validation, and holdout) on the main considerations: first, we want to identify if the 3 considered classification problems combined with an sets have specific characteristics that single them out with Information Gain feature selection [53]. The results respect to the rest of the data; second, the nature of show a performance decrease when classifying MCI the adopted classifier is intrinsically binary and therefore vs AD, this can be caused by the similarity of the they are expected to perform better. Indeed, we obtain two stages of dementia. Indeed, Mild Cognitive Impair- poorer performances when 3-classes are used (accuracy ment become increasingly prone to develop Alzheimer’s below 50%). or another type of dementia. On the other hand, Additionally, the overall classification accuracy of 50.6% an improvement of the HC vs CASE and HC vs on 100 different random permutations, for each clas- MCI classification tasks were observed reaching even sification problem and for each EEG signal processing more than 80% of accuracy. Notably, distinguishing technique, confirms the reliability of our classification MCI from healthy control cases can be useful to aid models and the absence of over-fitting when considering the prediction of the development of later stages of real classes. dementia. Fig. 3 Scatter plot of three example features (i.e., W_5_42, W_6_13, W_16_23) extracted from the C4.5 tree for HC (red points) vs MCI (blue points) subjects. The x-axis and y-axis represent the feature values for W_5_42 vs W_6_13 on the left, for W_5_42 vs W_16_23 on the right Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 8 of 10 In line with most previous studies of EEG classification Authors’ contributions MCDC, EW, GF conceived and designed the work. PBr, SD, AB acquired, [23–26], the electrodes mainly discriminant to classify are collected, and organized the data. GF, EW, AC implemented the T, O, F, Fp, whereas the wavebands more recurring are computational pipeline and analyzed the data. GF, EW, MCDC, GFel, PB alpha, theta, and delta. After all, an enhanced activity in interpreted the results. PBr, AB validated medical research. PB and GFel validated machine learning research. GF, EW, MCDC drafted the manuscript. the theta and delta wavebands, as well as a decreased All authors critically revised the manuscript, approved the final manuscript and activity in the alpha and beta ones. Indeed, during cogni- take responsibility for the accuracy and integrity of the work. tive impairment, beta waves (observed in the parietal and Ethics approval and consent to participate frontal region of the scalp) replace alpha waves, whereas The Ethical Committee of the IRCCS Centro Neurolesi “Bonino-Pulejo” theta waves are associated with decreased cognitive activ- approved the study after an informed consent to participate to the study was ities such as focusing and attention [68]. signed by the enrolled subjects (reference number 40/2013). To conclude, the results on the new and extended data Competing interests set are improved with respect to our previous study The authors declare that they have no competing interests. on fewer subjects [34]. Thanks to the Wavelet Trans- Publisher’s Note form we obtain promising results also with an enhanced Springer Nature remains neutral with regard to jurisdictional claims in number of HC subjects that are more challenging to published maps and institutional affiliations. discriminate. Author details Institute of Systems Analysis and Computer Science A. Ruberti (IASI), National Conclusions Research Council (CNR), Via dei Taurini 19, 00185 Rome, Italy. SysBio Centre In this work, we proposed an analysis procedure for EEG for Systems Biology, Rome, Italy. Department of Engineering, Uninettuno International University, Corso Vittorio Emanuele II 39, 00186 Rome, Italy. signals classification of samples affected by neurodegenar- IRCCS Centro Neurolesi “Bonino-Pulejo”, Contrada Casazza, SS113, 98124 ative diseases, i.e., Mild Cognitive Impairment (MCI) and Messina, Italy. Alzheimer Disease (AD), with respect to Healthy Control Received: 24 July 2017 Accepted: 22 May 2018 samples (HC). The analysis is based on a preprocess- ing phase followed by a feature extraction (Fast Fourier and Wavelet Analysis), and a classification procedure that References relies on the well-known supervised learning approach to 1. Prince M, Wimo A, Guerchet M, Ali G, Wu Y, Prina M. World Alzheimer Report 2015. The global impact of dementia. An analysis of prevalence, distinguish the AD, MCI, and HC classes. We tested our incidence, cost & trends; Alzheimer’s Disease International: London. procedure on EEG signals recorded on 109 human sam- London: Alzheimer’s Disease International (ADI); 2015. ples (23 HC, 37 MCI, 49 AD). Through the combination 2. Winblad B, Amouyel P, Andrieu S, Ballard C, Brayne C, Brodaty H, of the Wavelet-based signal analysis and the tree-based Cedazo-Minguez A, Dubois B, Edvardsson D, Feldman H, et al. Defeating alzheimer’s disease and other dementias: a priority for european science classifier C4.5, we effectively identified HC, MCI, and and society. Lancet Neurol. 2016;15(5):455. AD experimental samples with better accuracy than the 3. Grundman M, Petersen RC, Ferris SH, Thomas RG, Aisen PS, Bennett DA, spectral analysis with Fast Fourier Transform. Foster NL, Jack Jr CR, Galasko DR, Doody R, et al. Mild cognitive impairment can be distinguished from alzheimer disease and normal Finally, we plan to extend the analysis on a new cohort aging for clinical trials. Arch Neurol. 2004;61(1):59–66. of patients by further increasing the number of consid- 4. Brooker D, Fontaine JL, Evans S, Bray J, Saad K. Public health guidance to ered samples, in particular for the less represented classes, facilitate timely diagnosis of dementia: Alzheimer’s cooperative valuation in europe recommendations. Int J Geriatr Psychiatr. 2014;29(7):682–93. to apply more advanced artifact removal techniques [69], 5. Cedazo-Minguez A, Winblad B. Biomarkers for alzheimer’s disease and and to study the extracted classification models for identi- other forms of dementia: clinical needs, limitations and future aspects. fying further electrodes and wavebands that are related to Exp Gerontol. 2010;45(1):5–14. 6. Hampel H, Frank R, Broich K, Teipel SJ, Katz RG, Hardy J, Herholz K, the investigated diseases. Bokde AL, Jessen F, Hoessler YC, et al. Biomarkers for alzheimer’s disease: academic, industry and regulatory perspectives. Nat Rev Drug Discov. Abbreviations 2010;9(7):560. AD: Alzheimer’s disease; DWT: Discrete wavelet transform (DWT); EEG: 7. DeKosky ST, Marek K. Looking backward to move forward: early detection Electroencephalography; FFT: Fast fourier transform; FT: Fourier transform; HC: of neurodegenerative disorders. Science. 2003;302(5646):830–4. Healthy control; MCI: Mild cognitive impairment; WT: Wavelet transform 8. Jackson CE, Snyder PJ. Electroencephalography and event-related Acknowledgements potentials as biomarkers of mild cognitive impairment and mild EW and AC wish to thank Riccardo Torlone for proposing the topic during its Alzheimer’s disease. Cambridge: Elsevier; 2008. big data course. 9. Poil S-S, De Haan W, van der Flier WM, Mansvelder HD, Scheltens P, Linkenkaer-Hansen K. Integrative eeg biomarkers predict progression to Funding alzheimer’s disease at the mci stage. Front Aging Neurosci. 2013;5:58. The authors were partially supported the Nanomax flagship project 10. Jasper HH. The ten twenty electrode system of the international (B81J12006270001) and by the EPIGEN flagship project (PB.P01). These federation. Electroencephalogr Clin Neurophysiol. 1958;10:371–5. funding sources had no role in the design of the study and collection, analysis, 11. Elbert T, Lutzenberger W, Rockstroh B, Berg P, Cohen R. Physical aspects and interpretation of data, and in writing the manuscript. of the eeg in schizophrenics. Biol Psychiatry. 1992;32(7):595–606. 12. Davidson PR, Jones RD, Peiris MT. Eeg-based lapse detection with high Availability of data and materials temporal resolution. IEEE Trans Biomed Eng. 2007;54(5):832–9. The data sets analyzed during the current study are available upon request 13. DeKosky ST, Marek K. Looking backward to move forward: early detection from the IRCCS Centro Neurolesi “Bonino-Pulejo” (MCDC). of neurodegenerative disorders. Science. 2003;302(5646):830–4. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 9 of 10 14. Snyder SM, Hall JR, Cornwell SL, Falk JD. Addition of eeg improves 37. Powell G, Percival I. A spectral entropy method for distinguishing regular accuracy of a logistic model that uses neuropsychological and and irregular motion of hamiltonian systems. J Phys A Math Gen. cardiovascular factors to identify dementia and mci. Psychiatry Res. 1979;12(11):2053. 2011;186(1):97–102. 38. MATLAB: Version 7.10.0 (R2010a). Natick: The MathWorks Inc.; 2010. 15. Hampel H, Lista S, Teipel SJ, Garaci F, Nisticò R, Blennow K, Zetterberg 39. Tumari SM, Sudirman R, Ahmad A. Selection of a suitable wavelet for H, Bertram L, Duyckaerts C, Bakardjian H, et al. Perspective on future role cognitive memory using electroencephalograph signal. California: of biological markers in clinical therapy trials of alzheimer’s disease: a Scientific Research Publishing. Engineering. 2013;5(05):15. long-range point of view beyond 2020. Biochem Pharmacol. 2014;88(4): 40. Rosso O, Martin M, Figliola A, Keller K, Plastino A. Eeg analysis using 426–49. wavelet-based information tools. J Neurosci Methods. 2006;153(2): 16. Rossini PM, Rossi S, Babiloni C, Polich J. Clinical neurophysiology of 163–82. aging brain: from normal aging to neurodegeneration. Prog Neurobiol. 41. Hazarika N, Chen JZ, Tsoi AC, Sergejew A. Classification of eeg signals 2007;83(6):375–400. using the wavelet transform. In: Digital Signal Processing Proceedings, 17. Jeong J. Eeg dynamics in patients with alzheimer’s disease. Clin 1997. DSP 97., 1997 13th International Conference On, vol. 1. Santorini: Neurophysiol. 2004;115(7):1490–505. IEEE; 1997. p. 89–92. 18. Falk TH, Fraga FJ, Trambaiolli L, Anghinah R. Eeg amplitude modulation 42. Sanei S, Chambers JA. EEG Signal Processing. River Street, Hoboken: John analysis for semi-automated diagnosis of alzheimer’s disease. EURASIP J Wiley & Sons; 2013. Adv Signal Proc. 2012;2012(1):1–9. 43. Subha DP, Joseph PK, Acharya R, Lim CM. Eeg signal analysis: a survey. J 19. Dauwels J, Vialatte F, Cichocki A. Diagnosis of alzheimers disease from Med Syst. 2010;34(2):195–212. eeg signals: Where are we standing?. Curr Alzheimer Res. 2010;7(6): 44. Kumar PS, Arumuganathan R, Sivakumar K, Vimal C. Removal of ocular 487–505. artifacts in the eeg through wavelet transform without using an eog 20. Lehmann C, Koenig T, Jelic V, Prichep L, John RE, Wahlund L-O, Dodge reference channel. Int J Open Problems Compt Math. 2008;1(3):188–200. Y, Dierks T. Application and comparison of classification algorithms for 45. Daubechies I, et al. Ten Lectures on Wavelets vol. 61. Portland: SIAM, recognition of alzheimer’s disease in electrical brain activity (eeg). J Society for Industrial & Applied Mathematics; 1992. Neurosci Methods. 2007;161(2):342–50. 46. Swee E, Elangovan M. Applications of symlets for denoising and load 21. Dunkin JJ, Leuchter AF, Newton TF, Cook IA. Reduced eeg coherence in forecasting. In: Higher-Order Statistics, 1999. Proceedings of the IEEE dementia: state or trait marker?. Biol Psychiatry. 1994;35(11):870–9. Signal Processing Workshop On. Caesarea: IEEE; 1999. p. 165–169. 22. Giaquinto S, Nolfe G, Vitali S. Eeg changes induced by oxiracetam on 47. Subasi A. Eeg signal classification using wavelet feature extraction and a diazepam-medicated volunteers. Clin Neuropharmacol. 1986;9:79. mixture of expert model. Expert Syst Appl. 2007;32(4):1084–93. 23. Cibils D. Dementia and qeeg (alzheimer’s disease). Suppl Clin 48. Tan P, Steinbach M, Kumar V. Introduction to Data Mining. Boston: Neurophysiol. 2002;54:289–94. Addison Wesley; 2005. 24. Kowalski JW, Gawel M, Pfeffer A, Barcikowska M. The diagnostic value of 49. Weitschek E, Felici G, Bertolazzi P. Clinical Data Mining: Problems, Pitfalls eeg in alzheimer disease: correlation with the severity of mental and Solutions. Los Alamitos: IEEE Computer Society; 2013. p. 90–94. impairment. J Clin Neurophysiol. 2001;18(6):570–5. http://doi.org/10.1109/DEXA.2013.42. 25. Arenas A, Brenner R, Reynolds CF. Temporal slowing in the elderly 50. Quinlan JR. Improved use of continuous attributes in c4. 5. Journal of revisited. Am J EEG Technol. 1986;26:105–14. artificial intelligence research. 1996;4:77–90. 26. Coben LA, Danziger WL, Berg L. Frequency analysis of the resting awake 51. Hall M, Frank E, Holmes G, Pfahringer B, Reutemann P, Witten IH. The eeg in mild senile dementia of alzheimer type. Electroencephalogr Clin weka data mining software: an update. SIGKDD Explor Newsl. 2009;11(1): Neurophysiol. 1983;55(4):372–80. 10–18. http://doi.org/10.1145/1656274.1656278. 27. Locatelli T, Cursi M, Liberati D, Franceschi M, Comi G. Eeg coherence in 52. Bertolazzi P, Felici G, Festa P, Fiscon G, Weitschek E. Integer alzheimer’s disease. Electroencephalogr Clin Neurophysiol. 1998;106(3): programming models for feature selection: New extensions and a 229–37. randomized solution algorithm. Eur J Oper Res. 2016;250(2):389–99. 28. Besthorn C, Förstl H, Geiger-Kabisch C, Sattel H, Gasser T, 53. Hall MA, Smith LA. Practical feature subset selection for machine Schreiter-Gasser U. Eeg coherence in alzheimer disease. learning. In: In Proceedings of the 21st Australian Computer Science Electroencephalogr Clin Neurophysiol. 1994;90(3):242–5. Conference. New York: Springer; 1998. p. 181–91. 29. Dauwels J, Srinivasan K, Ramasubba Reddy M, Musha T, Vialatte F-B, 54. Cohen WW. Fast effective rule induction. In: In Proceedings of the Twelfth Latchoumane C, Jeong J, Cichocki A. Slowing and loss of complexity in International Conference on Machine Learning. Burlington: Morgan alzheimer’s eeg: two sides of the same coin?. Int J Alzheimer’s Dis. 2011;2011:. Kaufmann; 1995. p. 115–23. 30. Polat K, Günes¸ S. Classification of epileptiform eeg using a hybrid system 55. Cristianini N, Shawe-Taylor J. An Introduction to Support Vector based on decision tree classifier and fast fourier transform. Appl Math Machines and Other Kernel-based Learning Methods. Cambridge: Comput. 2007;187(2):1017–26. Cambridge University Press; 2000. 31. Akrami A, Solhjoo S, Motie-Nasrabadi A, Hashemi-Golpayegani M-R. 56. Bishop CM. Neural Networks for Pattern Recognition. Oxford: Oxford Eeg-based mental task classification: linear and nonlinear classification of University Press; 1995. movement imagery. In: Engineering in Medicine and Biology Society, 57. Trzepacz PT, Yu P, Sun J, Schuh K, Case M, Witte MM, Hochstetler H, 2005. IEEE-EMBS 2005. 27th Annual International Conference of The. Hake A, Initiative ADN, et al. Comparison of neuroimaging modalities for Shanghai: IEEE; 2006. p. 4626–9. the prediction of conversion from mild cognitive impairment to 32. Huang C, Wahlund L-O, Dierks T, Julin P, Winblad B, Jelic V. alzheimer’s dementia. Neurobiol Aging. 2014;35(1):143–51. Discrimination of alzheimer’s disease and mild cognitive impairment by 58. Previtali F, Bertolazzi P, Felici G, Weitschek E. A novel method and equivalent eeg sources: a cross-sectional and longitudinal study. Clin software for automatically classifying alzheimer’s disease patients by Neurophysiol. 2000;111(11):1961–7. magnetic resonance imaging analysis. Comput Methods Prog Biomed. 33. Mohammadi M, Al-Azab F, Raahemi B, Richards G, Jaworska N, Smith D, 2017;143:89–95. de la Salle S, Blier P, Knott V. Data mining eeg signals in depression for 59. Jackson CE, Snyder PJ. Electroencephalography and event-related their diagnostic value. BMC Med Inform Decis Mak. 2015;15(1):108. potentials as biomarkers of mild cognitive impairment and mild 34. Fiscon G, Weitschek E, Felici G, Bertolazzi P, De Salvo S, Bramanti P, alzheimer’s disease. Alzheimer’s and Dementia. 2008;4(1):137–43. De Cola MC. Alzheimer’s disease patients classification through eeg 60. Hampel H, Lista S, Teipel SJ, Garaci F, Nisticò R, Blennow K, Zetterberg signals processing. In: Computational Intelligence and Data Mining H, Bertram L, Duyckaerts C, Bakardjian H, et al. Perspective on future role (CIDM), 2014 IEEE Symposium On. Orlando: IEEE; 2014. p. 105–12. of biological markers in clinical therapy trials of alzheimer’s disease: a 35. Homan RW, Herman J, Purdy P. Cerebral location of international 10–20 long-range point of view beyond 2020. Biochem Pharmacol. 2014;88(4): system electrode placement. Electroencephalogr Clin Neurophysiol. 426–49. 1987;66(4):376–82. 61. Petrosian A, Prokhorov D, Lajara-Nanson W, Schiffer R. Recurrent neural 36. Adeli H, Zhou Z, Dadmehr N. Analysis of eeg records in an epileptic network-based approach for early recognition of alzheimer’s disease in patient using wavelet transform. J Neurosci Methods. 2003;123(1):69–87. eeg. Clin Neurophysiol. 2001;112(8):1378–87. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 10 of 10 62. Chowdhury RH, Reaz MB, Ali MABM, Bakar AA, Chellappan K, Chang TG. Surface electromyography signal processing and classification techniques. Sensors. 2013;13(9):12431–12466. 63. Al-Timemy AH, Bugmann G, Escudero J, Outram N. Classification of finger movements for the dexterous hand prosthesis control with surface electromyography. IEEE J Biomed Health Inform. 2013;17(3):608–18. 64. Mahajan K, Rajput SM. A comparative study of ann and svm for eeg classification. Int J Eng Res Technol. 2012;1:1–6. 65. Costantini G, Casali D, Todisco M. An svm based classification method for eeg signals. In: Proceedings of the 14th WSEAS International Conference on Circuits, Corfu Island, Greece, vol. 2224. New York: WSEAS World Scientific and Engineering Academy; 2010. 66. Wu T, Yang B, Sun H. Eeg classification based on artificial neural network in brain computer interface. In: Life System Modeling and Intelligent Computing. New York: Springer; 2010. p. 154–62. 67. Jeong J. Eeg dynamics in patients with alzheimer’s disease. Clin Neurophysiol. 2004;115(7):1490–505. 68. Pizzagalli DA. Electroencephalography and high-density electrophysiological source localization. Handb Psychophysiology. 2007;3:56–84. 69. Chen X, Liu A, Chen Q, Liu Y, Zou L, McKeown MJ. Simultaneous ocular and muscle artifact removal from eeg data by exploiting diverse statistics. Comput Biol Med. 2017. http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png BMC Medical Informatics and Decision Making Springer Journals

Combining EEG signal processing with supervised methods for Alzheimer’s patients classification

Free
10 pages

Loading next page...
 
/lp/springer_journal/combining-eeg-signal-processing-with-supervised-methods-for-alzheimer-4oQ5cF7Qh7
Publisher
BioMed Central
Copyright
Copyright © 2018 by The Author(s)
Subject
Medicine & Public Health; Health Informatics; Information Systems and Communication Service; Management of Computing and Information Systems
eISSN
1472-6947
D.O.I.
10.1186/s12911-018-0613-y
Publisher site
See Article on Publisher Site

Abstract

Background: Alzheimer’s Disease (AD) is a neurodegenaritive disorder characterized by a progressive dementia, for which actually no cure is known. An early detection of patients affected by AD can be obtained by analyzing their electroencephalography (EEG) signals, which show a reduction of the complexity, a perturbation of the synchrony, and a slowing down of the rhythms. Methods: In this work, we apply a procedure that exploits feature extraction and classification techniques to EEG signals, whose aim is to distinguish patient affected by AD from the ones affected by Mild Cognitive Impairment (MCI) and healthy control (HC) samples. Specifically, we perform a time-frequency analysis by applying both the Fourier and Wavelet Transforms on 109 samples belonging to AD, MCI, and HC classes. The classification procedure is designed with the following steps: (i) preprocessing of EEG signals; (ii) feature extraction by means of the Discrete Fourier and Wavelet Transforms; and (iii) classification with tree-based supervised methods. Results: By applying our procedure, we are able to extract reliable human-interpretable classification models that allow to automatically assign the patients into their belonging class. In particular, by exploiting a Wavelet feature extraction we achieve 83%, 92%, and 79% of accuracy when dealing with HC vs AD, HC vs MCI, and MCI vs AD classification problems, respectively. Conclusions: Finally, by comparing the classification performances with both feature extraction methods, we find out that Wavelets analysis outperforms Fourier. Hence, we suggest it in combination with supervised methods for automatic patients classification based on their EEG signals for aiding the medical diagnosis of dementia. Keywords: Alzheimer’s disease, Feature extraction, Electroencephalography signals, Classification Background dementiarepresentsoneofthemajorplaguefor themod- Dementia is a broad group of brain disorders leading to ern society. The most widespread cause of dementia is the a cognitive impairment because of a gradual dysfunction Alzheimer’s disease (AD), which involves serious mem- and death of brain cells. The World Alzheimer Report ory loss, cognitive impairment, and behavioural changes. 2015 has been estimated that 36 million people were liv- Thus, AD interferes with daily, social and professional ing with dementia in 2010, nearly doubling every 20 years functioning of patients, also affecting the daily life of their to 66 million by 2030 and to 115 million by 2050 [1]. families [2]. The intermediate stage between the normal Given the continuous growth of incidence of this illness, cognitive deficit due to aging and dementia is defined as Mild Cognitive Impairment (MCI). Several symptoms dis- tinguish MCI, but the loss of memory is a risk factor to *Correspondence: giulia.fiscon@iasi.cnr.it Giulia Fiscon and Emanuel Weitschek contributed equally to this work. develop AD [3]. In Europe, only 50% of the patients with Institute of Systems Analysis and Computer Science A. Ruberti (IASI), National dementia receive a diagnosis by a specialist centre, and Research Council (CNR), Via dei Taurini 19, 00185 Rome, Italy tests for dementia are carried out after the patient has SysBio Centre for Systems Biology, Rome, Italy Full list of author information is available at the end of the article already started showing symptoms and the disease has © The Author(s). 2018 Open Access This article is distributed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits unrestricted use, distribution, and reproduction in any medium, provided you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license, and indicate if changes were made. The Creative Commons Public Domain Dedication waiver (http://creativecommons.org/publicdomain/zero/1.0/) applies to the data made available in this article, unless otherwise stated. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 2 of 10 progressed [4]. Usually, the process for obtaining a clinical subjects affected by neurodegenerative diseases (e.g., AD) diagnosis for dementia of a patient is mainly based on the or other pathologies (e.g., epilepsy). delivery of a questionnaire in order to assess its cognitive Nevertheless, AD and MCI subjects are characterized abilities. However, a timely diagnosis would facilitate care, by a huge variability and thus discriminating artifacts and reduce the progression of the disease, and improve the patterns similarities to physiological brain activity still patient’s management to alleviate the burden. This might remain a crucial issue. In this regard, EEG signal pro- be achieved through a combination of diagnosis criteria cessing integrated with computational algorithms based and reliable biomarkers. on machine learning methods may contribute to a deeper In the past years, significant progresses have been comprehension of the disease and simplify the work of made to detect the early stages of dementia through bio- neurologists providing an additional tool to diagnose the chemical, genetic, neuroimaging, and neurophysiological stage of dementia [20, 30–33]. biomarkers such as Electroencephalography (EEG) [5– In this paper, we propose a procedure based on EEG- 9]. EEG provides the electrical activity of the brain by signal preprocessing and automatic classification with tracking the connectivity of neurons in the recording supervised learning methods, and its application to dis- sites of the scalp [10], processing it with milliseconds criminate subjects belonging to AD, or MCI, or HC precision. The condition of the brain physiology can be classes. This is an extension of a preliminary work [34]in inferred from the EEG signals recorded, and thus abnor- which we processed an EEG data set composed of 49 AD, malities can be identified through the detection of unusual 37 MCI and 14 healthy controls subjects (HC) by means frequency patterns [11]. Indeed, different rhythms with of a spectrum analysis based on the Fourier Transforma- diverse frequency bands describe the activity of the brain tion, and we automatically classified them with supervised and can be recorded by EEG. Among them, the main ones machine learning methods. Here, we have increased the are alpha (8-13 Hz, 30-50 μV amplitude), beta (13-30 Hz, number of HC subjects of the data set to 23 in order 5-30 μV amplitude), gamma (≥ 30 Hz), delta (0.5-4 Hz), to balance the number of samples for each category. We and theta (4-7 Hz, ≥20μV amplitude). have also improved the EEG-signal preprocessing and Although it is characterized by a lower spatial resolution spectrum analysis techniques through the application of than other neuroimaging techniques, EEG provides high the Wavelet Transform as an efficient method for noise temporal resolution [12]. Moreover, EEG is non-invasive, reduction and feature extraction, obtaining a more reliable ease and faster to use and able to differentiate severity of method to distinguish healthy from diseased subjects. dementia at a lower cost than other imaging devices [13, 14]. Thanks to its reduced costs EEG can be easily imple- Methods mented for population screening to detect pre-clinical We apply a procedure that encompasses the following biomarkers. steps summarized in Fig. 1: (i) data collection (subjects EEG signal analysis may provide useful indications of recruitment, EEG recording) and preprocessing, (ii) fea- the patterns of brain activity and predict the stages of ture extraction (Fast Fourier and Wavelet Analysis), and dementia [15, 16] because of its significant capacity to (iii) classification (supervised machine learning to distin- detect brain rhythm abnormalities, generally correlated guishtheAD,MCI,and HCclasses). with the severity of cognitive impairment [17]. In par- ticular, different clinical studies confirm EEG as suitable Data collection and preprocessing technique to early detect AD [18–20], due to the follow- Subject Recruitment The IRCCS Centro Neurolesi ing effects on EEG signals: reduction of the complexity, “Bonino-Pulejo” enrolled in 2012 and 2013 a total of 109 perturbation of the synchrony, and slowingdown of the subjects: 86 patients affected by dementia (AD, MCI) rhythms [19, 21, 22]. The slowing of the rhythms in the of which 37 men and 49 women, and 23 healthy con- EEG signals of subjects affected by AD can be explained trols samples (HC) of which 13 men and 10 women. The by a gain of the activity in the theta and delta frequency patients have been classified either in AD or MCI, tak- ranges, and a reduction of the activity in the alpha and ing into account the World Health Organization standard. beta frequency ranges [23–26]. The reduction of complex- Subjects capable of undergoing an electroencephalogram ity in the EEG temporal patterns can be explained by a and with a negative anamnesis for neurological comorbid modification of the neural network architecture observed disease have been included. Conversely, subjects under in subjects affected by AD [27, 28] due to loss of neu- pharmacological treatment that could change the activity rons and functional interaction alteration which make the of the brain have been excluded from the study. Patients ac tivity of the brain more predictable, more regular, and are mean aged 78.4 ± 6.4 and 74.1 ± 9.4 years, respectively simplerthaninhealthy controlsamples (HC) [29]. There- for AD and MCI, whereas the mean age of healthy con- fore, we can state that EEG signals related to healthy trols is 65.6 ± 7.9 years. Association between gender and controls subjects can be distinguished from those ones of etiological class (AD, MCI, HC) is not detected by the chi- Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 3 of 10 Fig. 1 Flowchart of the EEG signal analysis procedure square test (p-value > 0.05). Additionally, the difference preprocessed EEG signals, we apply the Fourier and the in terms of age between men and women is not statis- Wavelet Transform for estimating their spectrum [37]. tically significant according to the two-tailed Student’s T Firstly, the Fast Fourier Transform (FFT) is applied to Test (p-value > 0.05 for each class). Thus, the hypothe- each EEG signal of 180 seconds providing M Fourier Coef- sis of homogeneity for age and gender among etiological ficients for each electrode (M equal to 16). Hence, for each classes cannot be rejected. Table 1 provides an overview of sample we obtain 304 features (16 coefficients · 19 elec- the enrolled subjects that can be divided in three main eti- trodes) and we arrange them in a matrix with 109 rows ological classes: (i) patients with Alzheimer’s disease (AD), (referring to the samples) and 305 columns (304 referring (ii) patients with Mild Cognitive Impairment (MCI), and to the features, and one referring to the sample type). (iii) healthy control samples (CT). Secondly, the Discrete Wavelet Transform (DWT) is applied to each EEG signal of 180 seconds providing M Wavelet Coefficients for each electrode (M equal to 48). EEG recording We acquired multi-channel EEG sig- Hence, for each sample we obtain 912 features (48 coef- nals by using 19 electrodes, by setting their placement ficients · 19 electrodes) and we arrange them in a matrix according to the International 10-20 System [35], and by with 109 rows (referring to the samples) and 913 columns exploiting monopolar connections with earlobe electrode (912 referring to the features, and one referring to the landmark [10]. The brain activity of the subjects in rest- sample type). ing condition and closed eyes was measured in terms We provide in Table 2 a schematic representation of the of electrical potential (μV ). We recorded the EEG sig- matrices. nals by capturing 300 seconds with 256 or 1024 sampling The spectral analysis (Wavelet and Fourier) of the EEG frequency (Hz). signals has been performed by using the high level com- o˝ puting language provided by MATLAB R2014a [38]. Preprocessing For each signal we select the central 180 seconds (i.e., from 60 to 240 seconds) to avoid initial and Fourier analysis final EEG recording artifacts. Additionally, to normalize We apply the Fast Fourier Transform (FFT) to obtain the the sampling frequency we convert each signal to 256 Hz. spectrum of the EEG signals [37]. The FFT relies on the Discrete Fourier Transform (DFT) computed as follows: Feature extraction S−1 Extracting features from EEG signals in frequency domain X[ k] = x[ s] e [ s](1) hasbeenproventobeeffectiveforanalyzing theelectrical s=0 brain activity with computational models [31, 36]. Thus, in order to obtain a set of informative features from the Table 2 Schema of the matrix obtained after the feature extraction phase Table 1 Overview of the recruited subjects Sample Coefficient ··· Coefficient Sample type (1,1) (E,E·M) Sample Number of samples (%) Average age (std dev.) in years sample a ··· a HC 1 (1,1) (1,E·M) type Male Female Total Male Female Total sample a ··· a MCI 2 (2,1) (2,E·M) AD 20 (41%) 29 (59%) 49 78.6 (4.1) 78.2 (7.6) 78.4 (6.4) ··· ··· ··· ··· ··· MCI 17 (46%) 20 (54%) 37 75.7 (9.7) 72.7 (9.1) 74.1 (9.4) sample a ··· a AD N (N,1) (N,E·M) HC 13 (56%) 10 (44%) 23 68.1 (6.9) 62.3 (8.3) 65.6 (7.9) N = number of samples, M = number of coefficients, M + 1 = number of features, E Total 50 (46%) 59 (54%) 109 74.9 (8.2) 73.6 (9.9) 74.2 (9.1) = number of electrodes, a = element of the matrix Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 4 of 10 with s representing the s-th sample in the time domain; x In this work, we adopt the DWT in order to per- corresponding to the signal time series (s = 0, 1, 2, ··· , S− form the spectral analysis on the previously described 1); X referring to the representation of th frequency dataset (see section Data collection and preprocessing). domain for the time-series signal x; S representing the the The choice of a simple DWT stems from the need of whole number of samples of the signal x; k correspond- obtaining good performances over an arbitrary num- ing to k-th frequency component (k = 0, 1, ... , S − 1); ber of feature elements per channel and from the sam- jks2π − pling frequency of the input signals (256 Hz). We e [ s] = e referring to the k-th basis function. adopt two types of discrete wavelet families: Daubechies e [ s] is calculated simultaneously during the sampling (db) and Symlets (sym). Daubechies are compactly supported phase. Such a formula yields as output one complex num- orthonormal wavelets [45], while Symlets are symmetrical ber X[ k]for each k component. The output of the FFT wavelets proposed by Daubechies as modifications to the analysis are the Fourier Coefficients arranged in a matrix db family [46]. as showninTable 2. Given a single set of signals, each one is processed Wavelet analysis according to a feature extraction procedure composed of A more effective way for decomposing time and frequency two main phases: noise reduction and feature extraction. of the EEG signal, and for processing it is provided by the Firstly, we perform a noise reduction phase, where each Wavelet Transform (WT). WT is a time-frequency repre- EEG signal is decomposed in n levels (i.e., sub-bands) sentation of the signal, which is decomposed in different by applying a DWT (Symlets order 3 wavelet type). For windows of variable size, i.e., sub-bands. Conversely to every sub-band x an upper and lower threshold value is the FFT, the WT is able to catch the transient features of calculated as: the analyzed signal [39], i.e., it enables to keep both the Thr (x) = avg(x) + 1.5 · stdev(x) (3) temporal (spatial duration) and frequency information of up the signal. Indeed, WT allows to represent when tran- Thr (x) = avg(x) − 1.5 · stdev(x) (4) dwn sient events occur in the signal and with what intensity, as well as the time variations of the frequency contents [40]. The values of each sample s are then compared according Given a signal, WT decomposes it in simpler oscillating to the defined thresholds (3)and (4)and if s > Thr or i up functions called wavelets. A family of wavelets (ψ (t)) s < Thr then s is reduced as follows: s ∗ (Thr (x) − i i i up a,b dwn are derived from a unique mother wavelet ψ(t) by scal- Thr (x))/100. This step is performed in order to obtain dwn ing (dilating and contracting) and by shifting it to different an effective artifact reduction and to avoid possible infor- time positions [40, 41]. mation loss. The artifact removal phase operates on two levels of signal decomposition: level 5 and level 8. We 1 t − b choose these decomposition levels, because their ranges ψ (t) = √ ψ (2) a,b take into account the alpha, theta, beta, and delta band- |a| a widths, which are widely adopted for EEG analysis and In Eq. 2 t is the time variable, a ∈ R \ 0isthe scal- have been proven to be effective when dealing with ing parameter, and b ∈ R is the shifting parameter. The Alzheimer’s diseased patients (see “Background”section wavelets are localized in both time and frequency with for more details). The channel signal is then reconstructed respect to the sinusoidal waves of Fourier, which are bet- with the obtained values, which are given as input to the ter localized in frequency, but infinitely extended in time feature extraction phase. [40]. Additionally, the former are limited in band, i.e., they Secondly, for extracting the features, we adopt the are composed of a defined range of frequencies. Daubechies order 4 (db4) wavelet type with a sampling When dealing with digital signals that are frequency frequency of 256 Hz at decomposition level 5, which has band-limited, the continuous form of WT can be dis- been shown to guarantee a precise feature extraction in cretized according to the sampling theorem [42]. The the brainwaves frequencies [47], and we perform a large Discrete Wavelet Transform (DWT) allows to process dig- set of test with different parameters obtaining lower per- ital signals by keeping enough information in reasonable formances. The feature extraction phase extracts the fol- computational time. A relevant feature of the DWT is the lowing statistical features: mean, standard deviation, and combination with high and low pass filters, through which power spectral density of the wavelet coefficients. All the the signals can be processed to filter the high and low three feature types, representing the frequencies distribu- frequencies in order to compress and reduce the noise tion of the EEG signals, are calculated over the n epochs [43], e.g., hidden artifacts and background noise during of a channel-related signal. This phase makes use of the the EEG signals recording. Indeed, the WT is a well- decomposition levels obtained by applying the DWT to established signal representation and feature extraction the values produced during the noise reduction phase. technique for EEG processing [44]. Our method allows to apply an adaptive, threshold-based Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 5 of 10 noise/artifact removal to the main bandwidths (i.e., alpha, by computing standard statistical metrics, as accuracy, theta, beta, delta). We extract 16 features per channel precision, sensitivity, specificity, and F-measure and by when considering the combination of only two band- adopting a leave-one-out cross validation sampling proce- widths, i.e., alpha - theta or beta - delta, and 12 features per dure [48]. It is worth to note that the classification models channel when taking into account all the four brainwaves. can be adopted to classify new subjects whose diagnosis has not been already assessed and that could constitute The output of the DWT analysis are the Wavelet Coef- ficients arranged in a matrix as shown in Table 2. an independent validation set for further verifying the extracted models. Classification Finally, in order to prove the validity of the extracted We perform a supervised learning analysis in order to models we performed random permutations of class automatically classify the samples to their types (HC, membership for each classification problem and each sig- MCI, AD) by processing their associated features [48, 49]. nal processing technique (Fourier and Wavelet). We test if Supervised learning automatically assigns a sample into our procedure is able to extract meaningful classification a class by inferring a classification model from labeled models regardless of the class partition imposed on the data (training set). Our aim is to extract a human readable training set. This would be verified only in the presence of model specific for each type (HC, MCI, AD) of sample a marked overfitting behavior. containing a small subset of features, e.g., ‘if Wavelet > 0.3 and Wavelet < 0.6 then the sample can be classified Results as MCI”). This model can support clinicians to identify In this section, we provide the classification results rely- pivotal features related to the investigated neurodegen- ing on the features extracted with the Fourier and Wavelet erative disease and to diagnose new cases. In particular, Transforms applied on EEG signals of 180 s. Tables 3 and 4 we address the following classification problems: (i) HC present the results of the Decision Tree classifier consid- vs AD; (ii) HC vs MCI; (iii) MCI vs AD; (iv) HC vs ering the Fourier Transform and the Wavelet Transform, CASE (MCI+AD), where the CASE class is composed of respectively. AD joint to MCI samples in order to test the recogni- In particular, Table 3 presents the results of the Decision tion of the diseased patients with respect to the healthy Tree C4.5 classifier concerning the EEG signals with M = ones. Among the plethora of classification methods we 16 extracted Fourier Coefficients. We obtain 72%, 72%, use Decision Trees classifiers (i.e., C4.5 [50]), because they 80%, 75% of accuracy when dealing with HC vs AD, HC allow to handle noisy datasets and over-fitting with an vs MCI, MCI vs AD, HC vs CASE classification problems, ad-hoc parameters tuning. Additionally, Decision Trees respectively. provide the investigator with a compact, clear, and human Table 4 presents the performance of the Decision Tree readable classification model. C4.5 is an algorithm for the C4.5 classifier concerning the EEG signals processed with generation of decision trees used for classification. A deci- the Wavelet Transform. In all classification tasks, the fea- sion tree is a structure similar to a flow chart, where each ture extraction based on the Wavelet Transform achieves node denotes a test on an attribute, each branch repre- high classification performance in all metrics, obtaining sents a result of a test, and every leaf is labeled by a class. 83%, 92%, 79%, and 73% of accuracy when dealing with Indeed a node with outgoing edges is termed test node HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE classi- and the final nodes are the leaves. fication problems, respectively. In particular, the Wavelet In decision trees the classification model permits to pre- spectral analysis outperforms the Fourier analysis when dict the class of a sample based on its features. The dealing with EEG signals classification of HC vs AD, HC algorithm takes as input a set of classified data (train- vs MCI, and HC vs CASE. Conversely, for MCI vs AD both ing set) and the output is composed by leaf nodes, which define the belonging to a class attribute. Indeed, the path Table 3 Classification performance [%] by using M = 16 Fourier from the root to a specific leaf means that all the pred- Coefficients as features and a leave-one-out sampling with 72, 60, icates applied to the features of the sample are verified. 86, 109 folds for HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE, The validity of the three is verified on a set of labeled respectively samples (test set), but whose class is taken into account HC vs AD HC vs MCI MCI vs AD HC vs CASE only for verification of the class assignments. In this work, we use the J48 Java based implementation of C4.5 avail- Accuracy 72.2 71.7 80.2 74.7 able in the Weka package [51]. In addition, we performed Precision 71.1 78.9 80.2 74.0 a large battery of tests with other families of classifiers Sensitivity 72.2 71.7 80.2 74.7 (function-based, rule-based, naive-based, and Bayesian- Specificity 59.0 79.0 78.5 46.3 based), whose performances are not satisfying and hence F-measure 71.4 71.8 80.1 74.7 not reported. The classification performance is evaluated Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 6 of 10 Table 4 Classification performance [%] by using M = 48 Wavelet MCI vs AD, HC vs CASE permutated classification prob- Coefficients as features and a leave-one-out sampling with 72, 60, lems, respectively. Therefore, we obtain an overall average 86, 109 folds for HC vs AD, HC vs MCI, MCI vs AD, HC vs CASE, classification accuracy of 50.6%. respectively We also test other classification methods, such as function- HC vs AD HC vs MCI MCI vs AD HC vs CASE based, rule-based, naive-based, and Bayesian-based (e.g., Accuracy 83.3 91.7 79.1 73.4 RIPPER [54], SVM [55], the MultiLayer Perceptron [56]), whose performances are not satisfying and hence are Precision 83.3 91.8 79.3 74.7 not reported. For instance, we performed a large num- Sensitivity 83.3 91.7 79.1 73.4 ber of tests with SVMs by tuning the parameters and by Specificity 78.0 91.5 79.1 51.5 setting the complexity, the epsilon for round-off error, F-measure 83.3 91.7 79.1 74.0 the random seed, the tolerance to many different com- bination of values, but results were not above 65% of accuracy. Furthermore, we remark that the adopted C4.5 algo- signal processing methods lead to comparable classifica- rithm provides a classification model, a tree built on tion performances. Wavelet/Fourier Coefficients, from which the investiga- Another validation of the proposed procedure is based tor can derive the corresponding set of EEG electrodes. on the variation of the adopted sampling schema and by Here, the classification trees extracted in our performed applying also a feature selection step with the Information analyses mainly involve the electrodes T, O, F, Fp and the Gain (InfoGain) filter as evaluation measure followed by wavebands alpha, theta, and delta. the Ranker search method, which reached good perfor- Figure 2 depicts an example of such a tree and Fig. 3 mance in our previous study [52]. We perform both 10- shows the scatter plot of the features extracted from this fold cross validation sampling and holdout (90% training tree and related to the considered classes (MCI and HC). and 10% test percentage split) combined with Information Gain filter [53] obtaining again satisfying classification Discussion performance. We observe an improvement of the HC Although different neuroimaging techniques (e.g., Magnetic vs CASE and HC vs MCI classification tasks both for Resonance Imaging, Positron Emission Tomography) can Fourier and Wavelet transforms reaching even more than 80 and 90% of accuracy, respectively. Conversely, a perfor- be used for aiding the diagnosis of dementia providing mance decreasing is observed when distinguishing MCI quantitative data about the brain abnormalities [57, 58], EEG is non-invasive, besides being cheaper, simpler vs AD, probably due to the similarity of the two classes. and faster to use than other imaging devices [13, 14]. Classification results are detailed in Table 5. For this reason, automated EEG signal analysis plays For validating our results and the extracted classifica- an important role in detecting dementia in the early tion models we apply the procedure to data with random stages, as well as in classifying disease severity [59–61]. permutations of class labels. This validation test is per- Supervised learning is doubtless one of the most popular formed on 100 different random permutations for each methods to classify brain disorders with EEG [62–66]. classification problem and for each EEG signal process- Several studies compared the performance of classifi- ing technique (Fourier and Wavelet). In particular, when cation algorithms in terms of sensitivity and specificity, using the Fourier (Wavelet) transform we achieve 56.5% both for the early detection of dementia and for aiding (52.5%), 45.0% (50.5%), 49.5% (45.9%), and 54.5% (50.4%) clinical diagnosis. According to [47], our results show of average accuracy considering HC vs AD, HC vs MCI, that the Wavelet spectral analysis outperforms the Fourier analysis in discriminating EEG of health controls to ones of demented patients. Since EEG may exhibit normal Table 5 Classification performance (Accuracy [%]) by using 10-fold cross validation (CV) sampling and holdout (90% training frequency and may appear similar to normal aged con- and 10% test percentage split) for HC vs AD, HC vs MCI, MCI vs trol subjects during the earliest stages of dementia [67], AD, HC vs CASE, taking into account Wavelet (WT) and Fourier the higher accuracy of C4.5 with WT in distinguishing (FT) Coefficients as features between HC and MCI is a notable result. It is prob- Wavelet Fourier ably due to the fact that WT is suitable for nonsta- 10-fold CV Holdout 10-fold CV Holdout tionary signal like EEG that provide linear combination of the sum of wavelet coefficients and mother wavelet HC vs AD 76.4 71.4 80.6 85.7 with frequency and localization information. In this way, HC vs MCI 93.3 83.3 83.3 83.3 WT is able to detect the slowing of the alpha rhythm, MCI vs AD 66.3 88.9 66.7 77.2 which is more commonly found in intermediate stages HC vs CASE 81.7 81.8 84.4 90.9 of AD [67]. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 7 of 10 Fig. 2 C4.5 tree for HC vs MCI of size 7 with 4 leaves. Each path from the root to a leaf represents a classification rule. Each leaf is associated to a class and two numbers. The first number is the total number of instances recognized by the rule, while the second optional number represents how many ones (if any) are misclassified The choice of using three 2-class classification mod- Furthermore, we tested two other sampling schemes els instead of a single 3-class one is motivated by two (i.e., 10-fold cross validation, and holdout) on the main considerations: first, we want to identify if the 3 considered classification problems combined with an sets have specific characteristics that single them out with Information Gain feature selection [53]. The results respect to the rest of the data; second, the nature of show a performance decrease when classifying MCI the adopted classifier is intrinsically binary and therefore vs AD, this can be caused by the similarity of the they are expected to perform better. Indeed, we obtain two stages of dementia. Indeed, Mild Cognitive Impair- poorer performances when 3-classes are used (accuracy ment become increasingly prone to develop Alzheimer’s below 50%). or another type of dementia. On the other hand, Additionally, the overall classification accuracy of 50.6% an improvement of the HC vs CASE and HC vs on 100 different random permutations, for each clas- MCI classification tasks were observed reaching even sification problem and for each EEG signal processing more than 80% of accuracy. Notably, distinguishing technique, confirms the reliability of our classification MCI from healthy control cases can be useful to aid models and the absence of over-fitting when considering the prediction of the development of later stages of real classes. dementia. Fig. 3 Scatter plot of three example features (i.e., W_5_42, W_6_13, W_16_23) extracted from the C4.5 tree for HC (red points) vs MCI (blue points) subjects. The x-axis and y-axis represent the feature values for W_5_42 vs W_6_13 on the left, for W_5_42 vs W_16_23 on the right Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 8 of 10 In line with most previous studies of EEG classification Authors’ contributions MCDC, EW, GF conceived and designed the work. PBr, SD, AB acquired, [23–26], the electrodes mainly discriminant to classify are collected, and organized the data. GF, EW, AC implemented the T, O, F, Fp, whereas the wavebands more recurring are computational pipeline and analyzed the data. GF, EW, MCDC, GFel, PB alpha, theta, and delta. After all, an enhanced activity in interpreted the results. PBr, AB validated medical research. PB and GFel validated machine learning research. GF, EW, MCDC drafted the manuscript. the theta and delta wavebands, as well as a decreased All authors critically revised the manuscript, approved the final manuscript and activity in the alpha and beta ones. Indeed, during cogni- take responsibility for the accuracy and integrity of the work. tive impairment, beta waves (observed in the parietal and Ethics approval and consent to participate frontal region of the scalp) replace alpha waves, whereas The Ethical Committee of the IRCCS Centro Neurolesi “Bonino-Pulejo” theta waves are associated with decreased cognitive activ- approved the study after an informed consent to participate to the study was ities such as focusing and attention [68]. signed by the enrolled subjects (reference number 40/2013). To conclude, the results on the new and extended data Competing interests set are improved with respect to our previous study The authors declare that they have no competing interests. on fewer subjects [34]. Thanks to the Wavelet Trans- Publisher’s Note form we obtain promising results also with an enhanced Springer Nature remains neutral with regard to jurisdictional claims in number of HC subjects that are more challenging to published maps and institutional affiliations. discriminate. Author details Institute of Systems Analysis and Computer Science A. Ruberti (IASI), National Conclusions Research Council (CNR), Via dei Taurini 19, 00185 Rome, Italy. SysBio Centre In this work, we proposed an analysis procedure for EEG for Systems Biology, Rome, Italy. Department of Engineering, Uninettuno International University, Corso Vittorio Emanuele II 39, 00186 Rome, Italy. signals classification of samples affected by neurodegenar- IRCCS Centro Neurolesi “Bonino-Pulejo”, Contrada Casazza, SS113, 98124 ative diseases, i.e., Mild Cognitive Impairment (MCI) and Messina, Italy. Alzheimer Disease (AD), with respect to Healthy Control Received: 24 July 2017 Accepted: 22 May 2018 samples (HC). The analysis is based on a preprocess- ing phase followed by a feature extraction (Fast Fourier and Wavelet Analysis), and a classification procedure that References relies on the well-known supervised learning approach to 1. Prince M, Wimo A, Guerchet M, Ali G, Wu Y, Prina M. World Alzheimer Report 2015. The global impact of dementia. An analysis of prevalence, distinguish the AD, MCI, and HC classes. We tested our incidence, cost & trends; Alzheimer’s Disease International: London. procedure on EEG signals recorded on 109 human sam- London: Alzheimer’s Disease International (ADI); 2015. ples (23 HC, 37 MCI, 49 AD). Through the combination 2. Winblad B, Amouyel P, Andrieu S, Ballard C, Brayne C, Brodaty H, of the Wavelet-based signal analysis and the tree-based Cedazo-Minguez A, Dubois B, Edvardsson D, Feldman H, et al. Defeating alzheimer’s disease and other dementias: a priority for european science classifier C4.5, we effectively identified HC, MCI, and and society. Lancet Neurol. 2016;15(5):455. AD experimental samples with better accuracy than the 3. Grundman M, Petersen RC, Ferris SH, Thomas RG, Aisen PS, Bennett DA, spectral analysis with Fast Fourier Transform. Foster NL, Jack Jr CR, Galasko DR, Doody R, et al. Mild cognitive impairment can be distinguished from alzheimer disease and normal Finally, we plan to extend the analysis on a new cohort aging for clinical trials. Arch Neurol. 2004;61(1):59–66. of patients by further increasing the number of consid- 4. Brooker D, Fontaine JL, Evans S, Bray J, Saad K. Public health guidance to ered samples, in particular for the less represented classes, facilitate timely diagnosis of dementia: Alzheimer’s cooperative valuation in europe recommendations. Int J Geriatr Psychiatr. 2014;29(7):682–93. to apply more advanced artifact removal techniques [69], 5. Cedazo-Minguez A, Winblad B. Biomarkers for alzheimer’s disease and and to study the extracted classification models for identi- other forms of dementia: clinical needs, limitations and future aspects. fying further electrodes and wavebands that are related to Exp Gerontol. 2010;45(1):5–14. 6. Hampel H, Frank R, Broich K, Teipel SJ, Katz RG, Hardy J, Herholz K, the investigated diseases. Bokde AL, Jessen F, Hoessler YC, et al. Biomarkers for alzheimer’s disease: academic, industry and regulatory perspectives. Nat Rev Drug Discov. Abbreviations 2010;9(7):560. AD: Alzheimer’s disease; DWT: Discrete wavelet transform (DWT); EEG: 7. DeKosky ST, Marek K. Looking backward to move forward: early detection Electroencephalography; FFT: Fast fourier transform; FT: Fourier transform; HC: of neurodegenerative disorders. Science. 2003;302(5646):830–4. Healthy control; MCI: Mild cognitive impairment; WT: Wavelet transform 8. Jackson CE, Snyder PJ. Electroencephalography and event-related Acknowledgements potentials as biomarkers of mild cognitive impairment and mild EW and AC wish to thank Riccardo Torlone for proposing the topic during its Alzheimer’s disease. Cambridge: Elsevier; 2008. big data course. 9. Poil S-S, De Haan W, van der Flier WM, Mansvelder HD, Scheltens P, Linkenkaer-Hansen K. Integrative eeg biomarkers predict progression to Funding alzheimer’s disease at the mci stage. Front Aging Neurosci. 2013;5:58. The authors were partially supported the Nanomax flagship project 10. Jasper HH. The ten twenty electrode system of the international (B81J12006270001) and by the EPIGEN flagship project (PB.P01). These federation. Electroencephalogr Clin Neurophysiol. 1958;10:371–5. funding sources had no role in the design of the study and collection, analysis, 11. Elbert T, Lutzenberger W, Rockstroh B, Berg P, Cohen R. Physical aspects and interpretation of data, and in writing the manuscript. of the eeg in schizophrenics. Biol Psychiatry. 1992;32(7):595–606. 12. Davidson PR, Jones RD, Peiris MT. Eeg-based lapse detection with high Availability of data and materials temporal resolution. IEEE Trans Biomed Eng. 2007;54(5):832–9. The data sets analyzed during the current study are available upon request 13. DeKosky ST, Marek K. Looking backward to move forward: early detection from the IRCCS Centro Neurolesi “Bonino-Pulejo” (MCDC). of neurodegenerative disorders. Science. 2003;302(5646):830–4. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 9 of 10 14. Snyder SM, Hall JR, Cornwell SL, Falk JD. Addition of eeg improves 37. Powell G, Percival I. A spectral entropy method for distinguishing regular accuracy of a logistic model that uses neuropsychological and and irregular motion of hamiltonian systems. J Phys A Math Gen. cardiovascular factors to identify dementia and mci. Psychiatry Res. 1979;12(11):2053. 2011;186(1):97–102. 38. MATLAB: Version 7.10.0 (R2010a). Natick: The MathWorks Inc.; 2010. 15. Hampel H, Lista S, Teipel SJ, Garaci F, Nisticò R, Blennow K, Zetterberg 39. Tumari SM, Sudirman R, Ahmad A. Selection of a suitable wavelet for H, Bertram L, Duyckaerts C, Bakardjian H, et al. Perspective on future role cognitive memory using electroencephalograph signal. California: of biological markers in clinical therapy trials of alzheimer’s disease: a Scientific Research Publishing. Engineering. 2013;5(05):15. long-range point of view beyond 2020. Biochem Pharmacol. 2014;88(4): 40. Rosso O, Martin M, Figliola A, Keller K, Plastino A. Eeg analysis using 426–49. wavelet-based information tools. J Neurosci Methods. 2006;153(2): 16. Rossini PM, Rossi S, Babiloni C, Polich J. Clinical neurophysiology of 163–82. aging brain: from normal aging to neurodegeneration. Prog Neurobiol. 41. Hazarika N, Chen JZ, Tsoi AC, Sergejew A. Classification of eeg signals 2007;83(6):375–400. using the wavelet transform. In: Digital Signal Processing Proceedings, 17. Jeong J. Eeg dynamics in patients with alzheimer’s disease. Clin 1997. DSP 97., 1997 13th International Conference On, vol. 1. Santorini: Neurophysiol. 2004;115(7):1490–505. IEEE; 1997. p. 89–92. 18. Falk TH, Fraga FJ, Trambaiolli L, Anghinah R. Eeg amplitude modulation 42. Sanei S, Chambers JA. EEG Signal Processing. River Street, Hoboken: John analysis for semi-automated diagnosis of alzheimer’s disease. EURASIP J Wiley & Sons; 2013. Adv Signal Proc. 2012;2012(1):1–9. 43. Subha DP, Joseph PK, Acharya R, Lim CM. Eeg signal analysis: a survey. J 19. Dauwels J, Vialatte F, Cichocki A. Diagnosis of alzheimers disease from Med Syst. 2010;34(2):195–212. eeg signals: Where are we standing?. Curr Alzheimer Res. 2010;7(6): 44. Kumar PS, Arumuganathan R, Sivakumar K, Vimal C. Removal of ocular 487–505. artifacts in the eeg through wavelet transform without using an eog 20. Lehmann C, Koenig T, Jelic V, Prichep L, John RE, Wahlund L-O, Dodge reference channel. Int J Open Problems Compt Math. 2008;1(3):188–200. Y, Dierks T. Application and comparison of classification algorithms for 45. Daubechies I, et al. Ten Lectures on Wavelets vol. 61. Portland: SIAM, recognition of alzheimer’s disease in electrical brain activity (eeg). J Society for Industrial & Applied Mathematics; 1992. Neurosci Methods. 2007;161(2):342–50. 46. Swee E, Elangovan M. Applications of symlets for denoising and load 21. Dunkin JJ, Leuchter AF, Newton TF, Cook IA. Reduced eeg coherence in forecasting. In: Higher-Order Statistics, 1999. Proceedings of the IEEE dementia: state or trait marker?. Biol Psychiatry. 1994;35(11):870–9. Signal Processing Workshop On. Caesarea: IEEE; 1999. p. 165–169. 22. Giaquinto S, Nolfe G, Vitali S. Eeg changes induced by oxiracetam on 47. Subasi A. Eeg signal classification using wavelet feature extraction and a diazepam-medicated volunteers. Clin Neuropharmacol. 1986;9:79. mixture of expert model. Expert Syst Appl. 2007;32(4):1084–93. 23. Cibils D. Dementia and qeeg (alzheimer’s disease). Suppl Clin 48. Tan P, Steinbach M, Kumar V. Introduction to Data Mining. Boston: Neurophysiol. 2002;54:289–94. Addison Wesley; 2005. 24. Kowalski JW, Gawel M, Pfeffer A, Barcikowska M. The diagnostic value of 49. Weitschek E, Felici G, Bertolazzi P. Clinical Data Mining: Problems, Pitfalls eeg in alzheimer disease: correlation with the severity of mental and Solutions. Los Alamitos: IEEE Computer Society; 2013. p. 90–94. impairment. J Clin Neurophysiol. 2001;18(6):570–5. http://doi.org/10.1109/DEXA.2013.42. 25. Arenas A, Brenner R, Reynolds CF. Temporal slowing in the elderly 50. Quinlan JR. Improved use of continuous attributes in c4. 5. Journal of revisited. Am J EEG Technol. 1986;26:105–14. artificial intelligence research. 1996;4:77–90. 26. Coben LA, Danziger WL, Berg L. Frequency analysis of the resting awake 51. Hall M, Frank E, Holmes G, Pfahringer B, Reutemann P, Witten IH. The eeg in mild senile dementia of alzheimer type. Electroencephalogr Clin weka data mining software: an update. SIGKDD Explor Newsl. 2009;11(1): Neurophysiol. 1983;55(4):372–80. 10–18. http://doi.org/10.1145/1656274.1656278. 27. Locatelli T, Cursi M, Liberati D, Franceschi M, Comi G. Eeg coherence in 52. Bertolazzi P, Felici G, Festa P, Fiscon G, Weitschek E. Integer alzheimer’s disease. Electroencephalogr Clin Neurophysiol. 1998;106(3): programming models for feature selection: New extensions and a 229–37. randomized solution algorithm. Eur J Oper Res. 2016;250(2):389–99. 28. Besthorn C, Förstl H, Geiger-Kabisch C, Sattel H, Gasser T, 53. Hall MA, Smith LA. Practical feature subset selection for machine Schreiter-Gasser U. Eeg coherence in alzheimer disease. learning. In: In Proceedings of the 21st Australian Computer Science Electroencephalogr Clin Neurophysiol. 1994;90(3):242–5. Conference. New York: Springer; 1998. p. 181–91. 29. Dauwels J, Srinivasan K, Ramasubba Reddy M, Musha T, Vialatte F-B, 54. Cohen WW. Fast effective rule induction. In: In Proceedings of the Twelfth Latchoumane C, Jeong J, Cichocki A. Slowing and loss of complexity in International Conference on Machine Learning. Burlington: Morgan alzheimer’s eeg: two sides of the same coin?. Int J Alzheimer’s Dis. 2011;2011:. Kaufmann; 1995. p. 115–23. 30. Polat K, Günes¸ S. Classification of epileptiform eeg using a hybrid system 55. Cristianini N, Shawe-Taylor J. An Introduction to Support Vector based on decision tree classifier and fast fourier transform. Appl Math Machines and Other Kernel-based Learning Methods. Cambridge: Comput. 2007;187(2):1017–26. Cambridge University Press; 2000. 31. Akrami A, Solhjoo S, Motie-Nasrabadi A, Hashemi-Golpayegani M-R. 56. Bishop CM. Neural Networks for Pattern Recognition. Oxford: Oxford Eeg-based mental task classification: linear and nonlinear classification of University Press; 1995. movement imagery. In: Engineering in Medicine and Biology Society, 57. Trzepacz PT, Yu P, Sun J, Schuh K, Case M, Witte MM, Hochstetler H, 2005. IEEE-EMBS 2005. 27th Annual International Conference of The. Hake A, Initiative ADN, et al. Comparison of neuroimaging modalities for Shanghai: IEEE; 2006. p. 4626–9. the prediction of conversion from mild cognitive impairment to 32. Huang C, Wahlund L-O, Dierks T, Julin P, Winblad B, Jelic V. alzheimer’s dementia. Neurobiol Aging. 2014;35(1):143–51. Discrimination of alzheimer’s disease and mild cognitive impairment by 58. Previtali F, Bertolazzi P, Felici G, Weitschek E. A novel method and equivalent eeg sources: a cross-sectional and longitudinal study. Clin software for automatically classifying alzheimer’s disease patients by Neurophysiol. 2000;111(11):1961–7. magnetic resonance imaging analysis. Comput Methods Prog Biomed. 33. Mohammadi M, Al-Azab F, Raahemi B, Richards G, Jaworska N, Smith D, 2017;143:89–95. de la Salle S, Blier P, Knott V. Data mining eeg signals in depression for 59. Jackson CE, Snyder PJ. Electroencephalography and event-related their diagnostic value. BMC Med Inform Decis Mak. 2015;15(1):108. potentials as biomarkers of mild cognitive impairment and mild 34. Fiscon G, Weitschek E, Felici G, Bertolazzi P, De Salvo S, Bramanti P, alzheimer’s disease. Alzheimer’s and Dementia. 2008;4(1):137–43. De Cola MC. Alzheimer’s disease patients classification through eeg 60. Hampel H, Lista S, Teipel SJ, Garaci F, Nisticò R, Blennow K, Zetterberg signals processing. In: Computational Intelligence and Data Mining H, Bertram L, Duyckaerts C, Bakardjian H, et al. Perspective on future role (CIDM), 2014 IEEE Symposium On. Orlando: IEEE; 2014. p. 105–12. of biological markers in clinical therapy trials of alzheimer’s disease: a 35. Homan RW, Herman J, Purdy P. Cerebral location of international 10–20 long-range point of view beyond 2020. Biochem Pharmacol. 2014;88(4): system electrode placement. Electroencephalogr Clin Neurophysiol. 426–49. 1987;66(4):376–82. 61. Petrosian A, Prokhorov D, Lajara-Nanson W, Schiffer R. Recurrent neural 36. Adeli H, Zhou Z, Dadmehr N. Analysis of eeg records in an epileptic network-based approach for early recognition of alzheimer’s disease in patient using wavelet transform. J Neurosci Methods. 2003;123(1):69–87. eeg. Clin Neurophysiol. 2001;112(8):1378–87. Fiscon et al. BMC Medical Informatics and Decision Making (2018) 18:35 Page 10 of 10 62. Chowdhury RH, Reaz MB, Ali MABM, Bakar AA, Chellappan K, Chang TG. Surface electromyography signal processing and classification techniques. Sensors. 2013;13(9):12431–12466. 63. Al-Timemy AH, Bugmann G, Escudero J, Outram N. Classification of finger movements for the dexterous hand prosthesis control with surface electromyography. IEEE J Biomed Health Inform. 2013;17(3):608–18. 64. Mahajan K, Rajput SM. A comparative study of ann and svm for eeg classification. Int J Eng Res Technol. 2012;1:1–6. 65. Costantini G, Casali D, Todisco M. An svm based classification method for eeg signals. In: Proceedings of the 14th WSEAS International Conference on Circuits, Corfu Island, Greece, vol. 2224. New York: WSEAS World Scientific and Engineering Academy; 2010. 66. Wu T, Yang B, Sun H. Eeg classification based on artificial neural network in brain computer interface. In: Life System Modeling and Intelligent Computing. New York: Springer; 2010. p. 154–62. 67. Jeong J. Eeg dynamics in patients with alzheimer’s disease. Clin Neurophysiol. 2004;115(7):1490–505. 68. Pizzagalli DA. Electroencephalography and high-density electrophysiological source localization. Handb Psychophysiology. 2007;3:56–84. 69. Chen X, Liu A, Chen Q, Liu Y, Zou L, McKeown MJ. Simultaneous ocular and muscle artifact removal from eeg data by exploiting diverse statistics. Comput Biol Med. 2017.

Journal

BMC Medical Informatics and Decision MakingSpringer Journals

Published: May 31, 2018

References

You’re reading a free preview. Subscribe to read the entire article.


DeepDyve is your
personal research library

It’s your single place to instantly
discover and read the research
that matters to you.

Enjoy affordable access to
over 18 million articles from more than
15,000 peer-reviewed journals.

All for just $49/month

Explore the DeepDyve Library

Search

Query the DeepDyve database, plus search all of PubMed and Google Scholar seamlessly

Organize

Save any article or search result from DeepDyve, PubMed, and Google Scholar... all in one place.

Access

Get unlimited, online access to over 18 million full-text articles from more than 15,000 scientific journals.

Your journals are on DeepDyve

Read from thousands of the leading scholarly journals from SpringerNature, Elsevier, Wiley-Blackwell, Oxford University Press and more.

All the latest content is available, no embargo periods.

See the journals in your area

DeepDyve

Freelancer

DeepDyve

Pro

Price

FREE

$49/month
$360/year

Save searches from
Google Scholar,
PubMed

Create lists to
organize your research

Export lists, citations

Read DeepDyve articles

Abstract access only

Unlimited access to over
18 million full-text articles

Print

20 pages / month

PDF Discount

20% off