EEG-MLP-IRPSO: an interpretable, lightweight framework for early parkinson’s disease detection via adaptive feature selection and generative augmentation

Abstract Parkinson’s disease (PD) requires early, non-invasive diagnosis, but current methods are limited by invasiveness, cost, and interpretability. We introduce EEG-MLP-IRPSO, a novel framework synergizing multi-domain Electroencephalogram (EEG) features, a pioneering Intelligent Relative Particle Swarm Optimization (IRPSO) for optimal feature selection (FS), and an optimized Multilayer Perceptron (MLP) classifier. IRPSO, featuring adaptive intelligence and dimension learning optimization (DLO), overcomes conventional FS challenges in high-dimensional EEG data, while a hybrid strategy employing conditional tabular generative adversarial network (CTGAN) and synthetic minority oversampling technique (SMOTE) addresses data scarcity. Our optimized MLP achieved 96.84% accuracy, 97.87% precision, and 99.52% AUC-ROC, significantly outperforming baselines, with results robustly validated. SHapley Additive exPlanations (SHAP) analysis confirms clinical relevance and enhances interpretability. This work presents a robust, highly accurate, efficient, and interpretable EEG-based framework for PD diagnosis, demonstrating a promising step toward a new paradigm for neurodegenerative disease detection.