Answers are grounded in this paper and cite their sources.
Background/Objectives: Wearable models for Parkinson’s disease (PD) often use a single modality, overfit small cohorts, require complete inputs, give unfaithful explanations or suffer from subject leakage. We aimed to develop NeuroEdge-X, an explainable multi-expert model that fuses bilateral smartwatch motion from 11 standardized tasks with the 30-item non-motor symptom questionnaire of the same individuals. Methods: A questionnaire expert, an interpretable spectral tremor expert and a MiniRocket random-convolution motion expert are combined by a stacked meta-learner, with a motion-only meta-learner when the questionnaire is missing. The model was evaluated on 469 participants of the PhysioNet PADS dataset under repeated subject-independent cross-validation, and the motion expert was also tested on an external gait cohort recorded with foot-pressure sensors. Results: NeuroEdge-X reached 91.0% balanced accuracy (BA) and 97.0% AUROC for PD versus healthy controls and 75.0% BA and 82.7% AUROC for PD versus other movement disorders. Its gain over a fused logistic regression was small and not consistently significant; its practical advantages are that it retained 81.8% and 68.7% BA without the questionnaire and that its explanations are exact. The motion expert alone transferred to the external gait cohort (AUROC 83.5%). The explanations highlight tremor frequency, tremor regularity and action tasks; calibration, decision-curve and triage analyses support clinical interpretation. Profiled with PyTorch, ONNX Runtime and scikit-learn on one thread of a general-purpose CPU, the model processed a participant in 48 ms with 2.3 MB of working memory. Conclusions: Because the questionnaire was completed after diagnosis, these estimates may be optimistic for undiagnosed people. The complete framework has not been validated externally, and edge deployment was not verified on an ultra-low-power wearable microcontroller.
No takes yet. Share an insight, caveat, or question.
Khan et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: