Unimodal vs. Multimodal Data in Early-Onset Parkinson’s Disease: A Comparative Machine Learning Study

Abstract

Early-onset Parkinson’s disease  presents subtle and overlapping motor and non-motor symptoms, making early diagnosis challenging. This study compares unimodal and multimodal machine learning  frameworks for EOPD detection using the UCI Parkinson’s voice dataset. Four supervised algorithms Random Forest, Support Vector Machine, Logistic Regression, and Gradient Boosting Machine were implemented to evaluate model performance. Results revealed that the unimodal voice-based Random Forest model achieved the highest accuracy (94.9%), AUC (0.9485), and F1-score (0.9663), providing evidence that jitter, shimmer, and the harmonic-to-noise ratio are useful biomarkers for early diagnosis when obtained from speech recordings. Although multimodal fusion improved the average AUC (+10.7%) and F1-score (+7.1%), accuracy improvements were not statistically significant ( p > 0.05). These findings suggest that unimodal voice data can deliver strong diagnostic performance, while multimodal frameworks enhance robustness and model confidence. The study supports the integration of data-driven and explainable AI models for early, non-invasive ‘Parkinson’s disease’ diagnosis.