Self-Supervised Temporal Representation Learning for Parkinson’s Disease Detection Using 3D Gait Data
Download| Volume 7 Issue 1, 2026 | |
|---|---|
| Author(s): |
Abdul Khaliq Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, khaliq@iobm.edu.pk Muhammad Asghar Khan Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, muhammad.asghar@iobm.edu.pk Muhammad Abbas Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, dr.abbas@iobm.edu.pk Sana Alam* Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, sana.alam@iobm.edu.pk Qurrat Ul Ain Naiyar Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, qurratulain.naiyar@iobm.edu.pk Najmul Islam Department of Computer Science, College of Computer Science and Information Systems, Institute of Business Management, Karachi, Pakistan, najmul.islam@iobm.edu.pk |
| Abstract | Parkinson’s disease (PD) is a progressive neurodegenerative disorder characterized by motor impairments, including tremor, rigidity, and postural instability. Early detection and continuous assessment of PD remain challenging, particularly when clinical datasets contain limited labeled samples. Recent advances in machine learning have enabled automated analysis of biomechanical data; however, most existing approaches rely on supervised learning and substantial annotated datasets. This study proposes a self-supervised temporal representation learning framework for PD detection using three-dimensional (3D) gait data. The framework employs a bidirectional Long Short-Term Memory (BiLSTM) network with a masked sequence reconstruction task to learn temporal motion representations from gait sequences. The pretrained encoder is subsequently fine-tuned to classify gait patterns associated with PD. The proposed framework was evaluated on the CARE-PD dataset containing 3D movement data from individuals with different clinical conditions. The model achieved an accuracy of 78.8%, a receiver operating characteristic area under the curve (ROC-AUC) of 0.81, and a reported recall of 0.90 for participants with PD. The proposed approach achieved the highest accuracy among the evaluated LSTM, BiLSTM, CNN-BiLSTM, and Transformer baseline models. The main contribution of this study is the use of masked sequence reconstruction to pretrain a temporal gait representation before supervised PD classification, thereby reducing the dependence on disease-specific labels. The findings provide initial evidence that self-supervised temporal representation learning can be useful for gait-based PD detection under limited labeled-data conditions. |
| Keywords | Parkinson’s disease, self-supervised learning, 3D gait analysis, temporal representation learning, BiLSTM, masked sequence reconstruction, deep learning, gait-based detection. |
| Year | 2026 |
| Volume | 7 |
| Issue | 1 |
| Type | Research paper, manuscript, article |
| Recognized by | Higher Education Commission of Pakistan, HEC | Category | Journal Name | ILMA Journal of Technology & Software Management | Publisher Name | ILMA University | Jel Classification | -- | DOI | - | ISSN no (E, Electronic) | 2790-590X | ISSN no (P, Print) | 2709-2240 | Country | Pakistan | City | Karachi | Institution Type | University | Journal Type | Open Access | Manuscript Processing | Blind Peer Reviewed | Format | Paper Link | https://ijtsm.ilmauniversity.edu.pk/arc/Vol7/i1/pdf5.pdf | Page | 42-53 |