An Intelligent Machine Learning Framework for Multi-Class Detection of Neurological Disorders Using Physiological Signals

Abstract

Neurological and neurodevelopmental disorders often exhibit overlapping clinical manifestations, making early and accurate diagnosis a significant challenge in healthcare practice. Recent advances in physiological sensing technologies have enabled the continuous acquisition of health-related signals that can support intelligent disease screening systems. This study presents an intelligent machine learning framework for the multi-class detection of neurological disorders using multimodal physiological signals. The proposed framework utilizes heart rate, oxygen saturation (SpO₂), temperature, movement characteristics, infrared sensor measurements, and tri-axial accelerometer data to classify six neurological conditions, including Alzheimer face, Angelman syndrome, MECP2 syndrome, Pitt-Hopkins syndrome, Rett syndrome, and normal cases. To enhance predictive performance, a comprehensive preprocessing pipeline incorporating feature engineering, data normalization, Synthetic Minority Oversampling Technique (SMOTE), mutual information-based feature selection, and Principal Component Analysis (PCA) is employed. Multiple machine learning algorithms, including Logistic Regression, K-Nearest Neighbors, Support Vector Machines, Random Forest, Gradient Boosting, Extra Trees, XGBoost, LightGBM, CatBoost, Voting, and Stacking classifiers, are systematically evaluated. Experimental results demonstrate the superiority of ensemble learning approaches, with LightGBM achieving an accuracy of 99.60%, precision of 99.57%, recall of 99.63%, F1-score of 99.59%, Matthews Correlation Coefficient of 99.52%, and ROC-AUC of 99.99%. These findings highlight the effectiveness of multimodal physiological signal analysis and advanced machine learning techniques for reliable neurological disorder classification, providing a promising decision-support tool for early diagnosis and intelligent healthcare applications.