What Remains of Multimodal Deep Learning After Controlling Confounding and Subject Leakage? A Critical Evaluation in Smartphone Parkinson’s Detection

Reported AUROCs for smartphone-based Parkinson's detection often exceed 0.90. Three design choices inflate that number: modalities paired at random across unrelated datasets, PD participants who are on average 25.7 years older than controls, and the same person appearing on both sides of the split. We built a paired tri-modal cohort from mPower where one person completed walking, voice and tapping within a day, matched cases to controls on age and sex (age gap 25.7 to 0.4 years, 806 participants, 403 pairs), and grouped every split by participant. A random forest on twelve tapping features gets 0.751; the full deep multimodal transformer gets 0.726, with overlapping intervals. Only tapping mattered: gait alone 0.538, voice 0.583, and adding both moved the result by +0.001. JEPA pretraining was worth +0.012. The permutation null was 0.503. We also found that a widely used spiral benchmark has 102 images but only 28 subjects, with 25 shared between its train and test folders. Controlling all three problems, multimodal deep learning showed no advantage here.