Self-Supervised Masked Gait Modeling: An Explainable Low-Resource Learning Approach for Parkinson's Disease Detection
IEEE Access, cilt.14, ss.82890-82899, 2026 (SCI-Expanded, Scopus)
- Yayın Türü: Makale / Tam Makale
- Cilt numarası: 14
- Basım Tarihi: 2026
- Doi Numarası: 10.1109/access.2026.3697731
- Dergi Adı: IEEE Access
- Derginin Tarandığı İndeksler: Science Citation Index Expanded (SCI-EXPANDED), Scopus, Compendex, INSPEC, Directory of Open Access Journals
- Sayfa Sayıları: ss.82890-82899
- Anahtar Kelimeler: deep learning, explainable AI, gait analysis, masked autoencoders, Parkinson's disease, self-supervised learning
- Van Yüzüncü Yıl Üniversitesi Adresli: Evet
Özet
The application of deep learning to Vertical Ground Reaction Force (VGRF) signals presents considerable potential for the automated diagnosis of Parkinson's Disease (PD). However, the scarcity of available training data poses a significant challenge, as conventional supervised learning models are prone to overfitting and often depend on labor-intensive, hand-crafted features that necessitate substantial domain-specific expertise. To mitigate the challenge of limited data availability, this study proposes the Masked Gait Modeling (MGM) framework, a self-supervised learning paradigm designed to construct robust and generalizable data representations. The underlying principle of this approach is conceptually intuitive: segments of VGRF time-series data are randomly masked, and a Transformer encoder is trained to reconstruct the obscured portions. Through this reconstruction objective, the model acquires an intrinsic understanding of the biomechanical characteristics of human gait, entirely without reliance on disease-specific labels. Evaluation on the PhysioNet Parkinson's Disease Gait dataset demonstrates that MGM surpasses conventional supervised baselines, achieving an accuracy of 95.94% and a weighted F1 score of 0.96. The model exhibits exceptional sensitivity, correctly identifying 98% of PD cases. To facilitate clinical transparency, gradient-based saliency maps were utilized to visualize the model's decision-making process. Significantly, even without explicit supervision, the model inherently directs its attention to the heel-strike and toe-off phases. These are the exact components of the gait cycle routinely evaluated by neurologists for motor impairments. Such outcomes emphasize the twofold advantage of utilizing self-supervised pre-training for neurodegenerative gait analysis, demonstrating both enhanced diagnostic accuracy and fundamental clinical transparency.