ParkFuse: a trust-gated late-fusion framework with shortcut auditing for multimodal Parkinson’s disease detection

Introduction Multimodal deep learning has been proposed as a low-cost method for early screening of Parkinson's disease (PD), with most recent frameworks reporting accuracies exceeding 96% using k-fold cross-validation instead of a single fixed train/test partition. Methods We introduce ParkFuse, a trust-gated late-fusion framework that assigns each modality a matched representation-classical learned acoustic descriptors for voice, geometric stroke and tremor descriptors for handwriting, and residual convolutional networks for images-combined through soft-voting late fusion. Out-of-fold predictions are obtained via stratified, grouped k-fold cross-validation, and a quantitative shortcut audit tests whether classes are separable by trivial acquisition properties alone, yielding a trust gate that excludes leakage-suspect modalities from fusion. Results On a public 10-modality PD dataset, the audit reveals that imaging modalities achieve a near-optimal train-validation gap, whereas the electrocardiogram (ECG) folder contains files that are byte-identical across both classes-a signature of leakage rather than disease. In contrast, sustained-vowel voice recordings remain the strongest valid predictor (achieving out-of-fold AUCs of 0.98 and 0.96). A controlled ablation study shows that modality-matched features outperform a uniform image-style network on voice and drawing by as much as 0.43 in AUC. Discussion We argue that headline multimodal PD accuracies must always be reported alongside leakage diagnostics and that voice remains the most trustworthy modality for screening.