Swarm driven automatic feature selection and classification framework for parkinson voice data

Abstrak

Parkinson’s disease (PD) severely impairs motor and vocal functions, and early detection is crucial for effective intervention. Conventional diagnostic procedures remain subjective and time-consuming, highlighting the need for automated, data-driven approaches. This study aims to develop an intelligent and fully automated framework integrating Particle Swarm Optimization (PSO)–based feature selection with ensemble machine learning classifiers for PD detection using voice data. The proposed Swarm-Driven Automatic Feature Selection and Classification Framework (SAFSCF) automates data preprocessing, adaptive feature optimization, and classification within a unified pipeline. The framework was evaluated on the Parkinson’s Speech Dataset comprising 743 numerical features. Baseline models achieved accuracies of 0.7738 (Logistic Regression), 0.8651 (Random Forest), and 0.8690 (Gradient Boosting). After PSO optimization, the feature set was reduced by nearly 50% to 382 attributes, achieving a test accuracy of 0.8421 slightly higher than the full-feature model (0.8355). Convergence plots confirmed that PSO effectively minimized the fitness function while maintaining high classification stability. Feature importance analysis revealed that the most discriminative attributes were derived from log energy, Teager Kaiser energy operators (TKEO), MFCCs, Shimmer, and entropy-based features biomarkers known to reflect Parkinsonian speech degradation. These findings demonstrate that the proposed framework enhances computational efficiency and interpretability, offering a reproducible and scalable solution for non-invasive, voice-based PD diagnosis.