Lightweight pose detection for marathon sports based on improved Center Net algorithmChen, Wei; Wan, Shanqing
doi: 10.1007/s44163-026-00886-0pmid: N/A
The human pose detection in marathon sports faces challenges such as large motion amplitude, multiple limb occlusions, and limited computing resources. Traditional detection models are prone to accuracy degradation and response delay in practical applications. Therefore, the study introduces an attention mechanism module based on multi-scale channel weighting and deformable convolution to enhance feature expression ability. The detection head is designed to be lightweight through channel by channel convolution mechanism and Squeeze channel compression mechanism. Finally, a lightweight Center Net pose detection model that integrates multi-module optimization is proposed. The proposed model achieved F1 values of 92.83% and 94.37% on the Human 3.6 M Human Motion Capture Dataset-Running Subset and the AI Challenger Human Keypoint Detection Dataset-Running Category, respectively, with an average response time of less than 0.65 s, significantly better than that of the other three advanced models. The joint prediction error under different running poses was less than 4.5°, and the missing rate of key points in multi-light and camera shake scenes was controlled within 5%, with a frame rate of up to 35FPS. The model performs well in accuracy, robustness, and real-time performance, which is suitable for human pose recognition and analysis tasks in intelligent terminals and sports scenes.
Path planning for construction site inspection robots based on the hybrid optimization of improved APF and improved ACOHu, Yu
doi: 10.1007/s44163-026-01100-xpmid: N/A
In response to the dual challenges of complex static obstacles and frequent dynamic disturbances in the construction site environment, this paper proposes a hybrid path planning strategy that integrates global static planning and local dynamic re-planning. The research first adopts the grid method to model the environment. For global static programming, the deterministic annealing (DA) algorithm is introduced to control the temperature parameters, the potential function of the artificial potential field (APF) method is improved, and the free energy potential function is constructed, effectively solving the problem that traditional APF is prone to fall into local minima. For local dynamic programming, the algorithm adaptively divides the environment into multiple sub-regions, transforms the traversal problem into a traveling salesman problem, optimizes the traversal sequence using the simulated annealing algorithm, and adopts an improved ant colony algorithm to search for the shortest paths between sub-regions and to deal with dynamic obstacles, thereby forming a collaborative planning system. Simulation experiments show that the proposed hybrid algorithm significantly improves performance: in an ultra-complex static environment, its path planning success rate is as high as 81.3%, the path length is reduced by 11.1% and 14.7% respectively compared with the A* and Dijkstra algorithms, and the computing time is reduced by 27.4% and 37.9% respectively. In a high dynamic obstacle environment, the success rate of local re-planning remains at 89.0%. Under extremely high obstacle density, the algorithm can still maintain a path stability of 77.6% and an obstacle avoidance success rate of 80.2%. In the tests conducted in actual construction site scenarios, it has also demonstrated excellent adaptability and optimization effects. This study effectively integrates the global optimization ability of DA-APF and the local adaptive ability of Simulated Annealing-Ant Colony Optimization (SA-ACO) Algorithm, significantly improving the real-time performance, obstacle avoidance accuracy and overall robustness of path planning, providing a better solution for the autonomous navigation of construction site inspection robots.
Design of a RAG framework for cardiology EHR analysisDefilippo, Annamaria; Canino, Giovanni; Procopio, Nicola; Trapuzzano , M. D. Albino; Sorrentino, Sabato; Indolfi, M. D. Ciro; Vizza, Patrizia; Veltri, Pierangelo; Guzzi, Pietro Hiram
doi: 10.1007/s44163-026-01128-zpmid: N/A
Large Language Models (LLMs) and Retrieval-Augmented Generation (RAG) systems have significantly advanced data modelling capabilities and improved opportunities for extracting knowledge from vast and heterogeneous biomedical datasets. Recent research has increasingly focused on integrating LLMs with custom-designed RAGs to create systems capable of handling complex biomedical challenges, with a growing demand for more reliable and precise prediction mechanisms in health-related contexts. This study introduces CardioTRAP, an architecture specifically designed to manage biomedical data, with a primary focus on cardiology. The system employs advanced indexing techniques to enable efficient storage and retrieval by integrating deep learning models that generate contextual and clinically relevant insights. By adopting a hybrid approach that combines supervised and unsupervised learning methods, CardioTRAP ensures both high accuracy and scalability, supporting predictive analytics, patient risk stratification, and the discovery of novel biomarkers. Benchmarks and practical applications, evaluated through state-of-the-art metrics, underscore its ability to enhance the identification of critical clinical features. Finally, CardioTRAP demonstrates how the integration of data management and RAG systems can serve as a bridge between biomedical research and clinical practice.
Bayesian network modelling for predicting employment tendency and major matching of vocational education studentsXiao, Yimin; Li, Qingyun
doi: 10.1007/s44163-026-00952-7pmid: N/A
The growing demand for accurate career guidance in vocational education necessitates the development of advanced models to predict employment trends and match students with suitable majors. Effective prediction and alignment can significantly enhance career readiness and reduce skill–job mismatches. Traditional approaches often rely on static surveys, linear regression models, or expert assessments, which lack adaptability, fail to capture probabilistic dependencies, and provide limited predictive accuracy. These shortcomings result in biased recommendations and reduced reliability in career planning. To address these limitations, this study introduces the Bayesian Employment and Major Matching (BEN-EM) framework, which leverages Bayesian network modelling to capture probabilistic relationships among academic performance, skill sets, career interests, and labour market demand. BEN-EM dynamically integrates multidimensional data to provide robust predictions under uncertainty. The proposed method can be applied by vocational education institutions, policymakers, and career counsellors to identify optimal major–career pathways, support evidence-based decision-making, and improve alignment between student skills and employer requirements. Experimental validation demonstrates that BEN-EM achieves higher predictive accuracy (0.75), greater adaptability (> 2.5), and lower mismatch rates than conventional methods. In addition, the framework attains a confidence score above 0.75, coverage of career options by 65%, Skill Alignment Score (0.80), and decision support effectiveness score (0.75–0.80), thereby enhancing vocational education outcomes and employment success.
Research on English text scoring technology based on deep learning in English teachingLi, Ailing
doi: 10.1007/s44163-025-00790-zpmid: N/A
Automated grading of essays has a considerable contribution to improving the efficiency of English education by reducing human bias and scalability in assessment. The paper presents a new hybrid scoring model using Bi LSTM, Transformer encoder and Light GBM to overcome the weaknesses inherent in conventional rule-based and machine learning methods. The paper utilizes the Deep Essays Dataset provided by Kaggle with diverse student populations with human-annotated scores. The approach entailed deep text preprocessing, shallow and semantic feature extraction and context representation fusion assisted by Bi LSTM and Transformer layers. These features are used as input to a Light GBM regressor for end score prediction. Experiments with large-scale data show that the new model outperforms all the current methods substantially on RMSE, MAE and Pearson correlation metrics. Moreover, the evaluation metrics like accuracy, precision, recall and F1-score are always more than 99%, which is a witness to the stability of the model. The graphical analysis using ROC and PR plots also serves as a witness to the robustness of the system. The findings prove the effectiveness of blending deep neural networks and gradient boosting to detect linguistic and semantic patterns in essays. This system represents a promising innovation in automated scoring technology, to lead to fair, reliable and scalable assessment in English learning.
Simulation-based deep learning for IoT-oriented music improvisation optimizationLiu, Meng
doi: 10.1007/s44163-026-01064-ypmid: N/A
Real-time music improvisation demands continuous audio effect adjustment, yet human control is limited by latency and cognitive load. Conventional model- or rule-based controllers struggle with cross-modal data fusion, often producing suboptimal trajectories. A dynamic architecture addresses this by analyzing multimodal cues. This research proposes a Music Effect Control with Chaotic Wing Suit Flying Search optimized-Seq2Seq (MEC-CWFSO-SeqNet) model, integrating attention-enhanced sequence modeling with a multi-layered MEC-CWFSO engine for adaptive hyperparameter tuning and output refinement. Multimodal inputs, including audio signals and motion sensor data, are processed using an encoder-decoder architecture. During training, the MEC-CWFSO-SeqNet algorithm iteratively adjusts latent embeddings to enhance predictive performance. A diverse dataset of expressive musical improvisations, incorporating instrumental and sensor-enhanced performance, is employed with high temporal resolution to ensure accurate multimodal alignment. Data augmentation techniques such as time stretching, pitch shifting, and Z-score normalization are applied to create balanced, key-agnostic sequences. Chroma features extracted via STFT and CQT were combined with performance and sensor data into 256-dimensional vectors. Transformer-based encoder layers model cross-domain relationships, while the attention-driven decoder predicts effect parameters. CWFSO dynamically optimizes decoder behavior and loss weights through chaotic search. MEC-CWFSO-SeqNet achieved improved performance with PF (+ 0.05), SPD (-0.31), ADP (-0.04), NDG (0.0033), PCG (0.017), and an accuracy of 98.8% compared to existing methods. Simulation-based measurements indicate an average latency of 46 ms per frame on Intel Core i9 CPU, below the ~ 100 ms perceptual threshold for live musical improvisation, demonstrating that attention-based sequence learning with chaotic meta-heuristic search enables a robust, timely expressive effect controller.
Image and metadata-driven personality inference for career recommendation: a social media-based AI framework for adolescentsIsmail, Heba; Alhefeiti, Maryam; Khalil, Ashraf
doi: 10.1007/s44163-026-00981-2pmid: N/A
This study presents a novel AI-based framework that leverages Instagram image and metadata analysis to infer Big Five personality traits and deliver personalized career recommendations for high school students in the UAE. Addressing the limitations of traditional recommender systems that rely on self-reported questionnaires or text, the proposed approach uses multimodal visual features—including profile metrics, HSV color patterns, semantic image labels, and texture analysis—to enable a non-intrusive, scalable personalization method. A pilot study involving data from 30 student accounts served as a proof of concept. Correlation analysis identified profile and HSV features as the most predictive, and four machine learning models were trained, with Logistic Regression achieving 97% accuracy (AUC 0.97) in personality prediction. The inferred traits were mapped to academic majors using a stereotype-based recommender system, achieving 90% alignment with student preferences as measured by the Electronic Emirati Scale for Professional Inclinations (EESPI). Findings demonstrate the feasibility and promise of integrating AI-driven image analysis with personality-aware recommendation. This work contributes to emerging trends in non-verbal, visual data-based personalization, particularly in educational and career guidance domains.
Why AI proficiency matters: reframing AI assimilation, agility, and performance in women-led venturesImtyaz, Faiza; Yameen, Mohammad
doi: 10.1007/s44163-026-01026-4pmid: N/A
Artificial intelligence (AI) and other advanced technological developments are becoming more widely recognised as essential drivers for increasing efficiency because they have the ability to completely transform almost every activity both inside and outside of an organisation. Nevertheless, the existing literature remains limited in its empirical investigations of how AI integration can enhance organisational outcomes, including entrepreneurial agility, customer agility, and entrepreneurial performance. Building upon the framework of technological dominance (TTD) and the dynamic capability view (DCV), this research examines the ramifications of AI assimilation (AIASS) on entrepreneurial performance (EP). Subsequently, it evaluates the role of entrepreneurial agility (EA) and customer agility (CA) as a mediator on the AIASS-EP link. Furthermore, the research probed the moderating influence of AI proficiency (AIP). This research employed a cross-sectional methodology, utilising 358 valid responses from Indian women entrepreneurs to validate the suggested framework. The results substantiate that AIASS is a significant determinant of EP, EA, and CA, having more potent impacts on CA. Additionally, EA and CA were identified as partly complementing mediators in the AIASS and EP link. Moreover, AIP was found to positively moderate the association between AIASS and EA. These results improve the TTD and DCV frameworks and offer practitioners important new information. These findings' ramifications are examined, emphasizing their applicability to both practice and research. Acknowledging some of the study's inherent limitations opens the door for further investigation.
A grid search ensemble framework for analyzing agricultural productivity based on agro-climatic factorsBhatnagar, Priya; Saxena, Abhishek
doi: 10.1007/s44163-026-01051-3pmid: N/A
Agricultural productivity is influenced by various factors, including climatic conditions, soil properties, and environmental factors. Traditional methods of analyzing the agro-climatic factors are based on manual and visual observation. It often fails to capture the intricate interactions that affect the sustainable growth of crops. Moreover, the efficacy of existing machine learning-based methods depends on the tuning of hyperparameters and requires a high level of human intervention. To overcome these issues, a grid search optimized ensemble framework has been proposed to analyze the agro-climatic factors affecting agriculture productivity. The ensemble framework integrates the random forest, gradient boosting, adaptive boosting, K-nearest neighbor, support vector regressor, and extreme gradient boosting in a single framework. These base models are combined using a stacking regressor with linear regression as a meta-model. Additionally, to refine the efficacy of the proposed ensemble framework, grid search is employed to tune the hyperparameters of these base models. To accomplish this, mean squared error is used as an objective fun2ction. Besides this, a Gradio-based graphical user interface is developed for real-time agro-climatic factor assessment. The interface serves as a farming assistance tool and provides an analysis of farming suitability based on the agro-climatic factors. The performance of the proposed method is judged on self-acquired agro-climatic data. The comparative quantitative assessment illustrates that the proposed method outperforms the existing methodologies for the custom agro-climatic data. The proposed method has a performance gain in terms of R² over KNN, RF, XG-Boost, G-Boost, AdaBoost, and SVR with values of + 4.83%, + 4.66%, + 1.78%, + 1.16%, + 1.37%, and + 1.77%. The feature importance analysis of each agro-climatic factor is also evaluated. The quantitative observation reveals that the minimum temperature has the highest impact on crop productivity.