Comparative Machine Learning Analysis of Acoustic and Drawing-Based Motor Biomarkers for Parkinson’s Disease Screening

Abstract

Voice and hand-drawing tests are low-cost, non-invasive behavioral probes of Parkinson’s disease, yet their relative reliability is rarely examined with attention to test-set size and class balance. This study compares acoustic and drawing-based motor biomarkers using the evaluations recorded for a portable screening prototype: 200 voice recordings described by 22 acoustic measures, 200 spiral drawings, and 200 wave drawings, each analyzed with 8 to 10 classifiers. Gradient-boosted tree models performed best: XGBoost for voice (90.0% accuracy), and CatBoost for spiral (87.5%) and wave drawings (89.5%). Confusion matrices recovered from the original evaluation show that the voice test set contained 30 recordings (24 patients, 6 healthy) and the spiral test set only 8 drawings (6 patients, 2 healthy). Against majority-class baselines of 80% and 75%, the best models were not significantly better than always predicting Parkinson’s disease (exact binomial p = 0.12 and 0.37), and 95% Wilson intervals were wide (voice 74.4–96.5%; spiral 52.9–97.8%). Class-specific analysis revealed different error profiles: the best voice model had a sensitivity of 95.8% but a specificity of 66.7%, whereas every spiral model avoided false alarms but missed patients, and linear and kernel models collapsed to predicting a single class. Feature-importance and voice–drawing fusion analyses could not be performed because raw data and participant links were unavailable, and a protocol for them is specified. Neither modality can yet be declared more reliable; about 140 test cases per modality are needed to estimate a 90% accuracy to within ±5 percentage points.