A Soft Set-Theoretic Framework for the Prediction and Personalization of Parkinson's Disease Management in Under-Resourced Settings

This study describes a framework utilizing fuzzy soft sets for the automatic classification of the severity of Parkinson's disease from voice telemonitoring data. The framework achieved an accuracy of 78.4% for the classification of mild, moderate, and severe disease while providing greater explainability than black box models. The analysis was based on 5875 voice recordings from 42 patients. Fuzzy membership functions were used to model the gradual voice deterioration and the correlation-based weighting was used to clinically informative features passes (shimmer, HNR: r = 0.45 - 0.51). Confidence scoring calibrated to the classifiers allowed the implementation of risk-stratified workflows, achieving 91.4% accuracy on high confidence predictions. The fuzzy method was 4.2% better than crisp soft sets s (p < 0.001). The framework also demonstrated the ability to handle naturally occurring patterns of missing data and the unexplainable classification made a case for the fuzzy soft sets to be used preferentially for clinical workflows.