Parkinson’s Disease Classification Using SHAP, LIME, and Fisher Score with XGBoost and White-Box Machine Learning Models: An Explainable AI Perspective

Abstract

Parkinson’s disease (PD) is a chronic neurological disorder that progressively impairs motor functions,often characterized by tremors, rigidity, and bradykinesia. Early diagnosis plays a vital role inimproving patient outcomes. This study applies machine learning (ML) classifiers for the accuratedetection of PD using a publicly available dataset comprising 195 voice recordings with 24 real-valuedspeech attributes. The preprocessing phase involved normalization and labeling, where the statusattribute encoded healthy (0) and PD (1) subjects. The models are evaluated using 10-fold crossvalidation to ensure robust performance estimation. The study evaluated a diverse range of machinelearning classifiers, encompassing interpretable white-box models along with complex black-boxmodels. Explainable Artificial Intelligence (XAI) techniques, including Fisher Score, SHAP, and LIME,are employed to interpret and rank feature importance, enhancing model transparency. Among allclassifiers, XGBoost achieved the best results, with an accuracy of 94.87%, F1-score of 96.97%, andROC-AUC of 0.91. The findings highlight that integrating XAI methods with ML not only yields highclassification accuracy but also provides interpretable insights essential for clinical decision support inPD diagnosis