Multi-Scale Adaptive Temporal Network for Parkinson’s Motor Symptom Fluctuation Detection Using Coarsened 2D Skeleton Pose Data

Background: Parkinson’s Disease (PD) is a neurological condition characterized by motor symptoms that fluctuate throughout the day depending on medication. Continuous and objective monitoring is essential, but conventional clinical assessments are episodic and subjective, while wearable and video-based solutions may raise privacy concerns. This study aims to develop a real-time, privacy-preserving deep learning framework that utilizes 2D skeleton pose data to simultaneously classify medication states (ON or OFF) and continuously estimate motor symptom severity. Methods: To enable privacy-preserving and real-time monitoring of Parkinson’s motor fluctuations, a Multi-Scale Temporal Attention-Transformer Network (MS-TATNet) was developed based on 2D skeleton pose data collected from the REal-world Mobility Activities in Parkinson’s disease dataset (REMAP) dataset. The MS-TATNet captures complex, variable, and multi-scale temporal dynamics of PD motor symptoms through a multi-scale temporal convolutional network, scaled dot-product attention mechanism, stacked transformer encoder blocks with a multi-head self-attention mechanism, temporal pooling layer, softmax classifier, and regression layer. Results: The experimental results demonstrate that the MS-TATNet achieved 99.63% accuracy, 99.50% recall, 99.33% specificity, and 99.67% F1-score for medication state classification. For continuous severity estimation, the predicted scores showed a Pearson correlation coefficient of 0.97 with clinical assessments. Conclusion: Thus, this work highlights the MS-TATNet’s potential for scalable, privacy-preserving remote monitoring of PD.