ParkinsonFusionNet: An Explainable Multimodal Deep Learning Framework for Parkinson's Disease Detection

Abstract

Parkinson's Disease (PD) impacts approximately 10 million people worldwide, but existing diagnostic approaches like neurological examination and neuroimaging usually only identify the disease after 60-80% of dopaminergic neurons have been lost, creating a need for automated, biomarker-driven screening tools. Here we propose to bridge this gap by presenting ParkinsonFusionNet, a multimodal deep learning framework that fuses three clinically validated non-invasive modalities: gait signals, voice biomarkers, and hand-tremor features. Feature representations are extracted by modality-specific encoders (CNN, RNN and transfer-learning architectures) and then fused and classified by a hard-voting ensemble of gradient-boosting models (XGBoost, CatBoost, LightGBM, Random Forest). Our method achieves a maximum accuracy of 98.31% and an AUC of 1.0000 on a 294-sample 42-feature multimodal dataset, outperforming all unimodal baselines. Feature-importance analysis indicates that hand-tremor amplitude, gait velocity, and voice Harmonic-to-Noise Ratio are the most diagnostically informative biomarkers, and that all three modalities contribute meaningfully. These results confirm that cross-modal fusion significantly outperforms diagnosis based on a single modality, and introduce a computationally efficient and clinically transparent framework suitable for scalable early PD screening in real-world clinical settings.