A Comparative Analysis of Machine Learning Techniques and Explainable AI on Voice Biomarkers for Effective Parkinson’s Disease Prediction

Parkinson's disease (PD) is a neurological movement disorder that remains difficult to diagnose, although it affects millions globally. Early diagnosis can lead to more effective and improved patient outcomes. Diagnosis through traditional methods is subjective and often lacks transparency, raising concerns about reliability. In this study, the CRISP-DM framework was applied to compare eight ML algorithms, including Random Forest and Support Vector Machine (SVM). Recursive Feature Elimination (RFE) was used to preprocess, balance, refine the data and find the eight most predictive vocal features. With 195 recordings coming from the UCI Parkinson’s Speech Dataset, which contains voice measurements from 31 individuals (23 with PD and 8 healthy controls), Random Forest (Entropy) had the best performance (F₁ = 96.6%, ROC AUC = 0.98). Explainable AI tools (SHAP and LIME) were integrated, allowing both global and instance-level understanding of model predictions thereby identifying measures of pitch variability (MDVP: RAP, spread1, PPE) as key predictors of PD. This research contributes to the practical deployment of reliable, transparent PD prediction tools in real-world medical settings, supporting early diagnosis and improved patient care. This raises the issue of the urgent need to detect PD early among Africa's aging populations to help protect the cultural heritage contained in the voices of the elders. this research contributes to the practical deployment of reliable, transparent PD prediction tools in real-world medical settings, supporting early diagnosis and improved patient care.Future work should embark on validating these findings over much more varied cohorts, integrating additional data modalities (e.g., gait, imaging), and enhancing model robustness. Real-time speech analysis-based tools, in the end, will allow remote screening, early intervention, and tailored care.