A Multi-Feature Deep Learning Framework for Assessing Leg Agility in Parkinson’s Disease

As global populations age, the prevalence of Parkinson’s disease (PD), a chronic degenerative condition affecting the central nervous system, continues to rise. Traditional assessment methods for PD often face limitations due to cost and complexity, hindering early detection and timely treatment adjustments. This study introduces an IoT-enabled dynamic assessment system for evaluating leg agility in PD patients, integrating computer vision, deep learning, and distributed edge-cloud computing. The proposed method addresses variability in body shape and appearance by normalizing skeletal data and centering the coordinate system. Positions of ankle and knee joints are determined through a human posture algorithm, normalized according to individual height, and analyzed using the Gramian angular field (GAF). The study explores the effectiveness of recurrent neural networks for time-series signal prediction and convolutional neural networks for GAF image analysis. A transformer encoder model enhanced with positional encoding and a multi-head self-attention mechanism is employed for sequence analysis. The model incorporates a feed-forward layer, layer normalization, and skip connections to ensure training stability. A residual CNN, based on channel fusion, is utilized to learn GAF image features effectively. Lastly, a multilayer perceptron (MLP) is used for feature fusion, integrating feature vectors from the self-attention encoder and GAF image, offering a robust strategy for diverse feature integration. This IoT-AI integration advances personalized tele-rehabilitation, enabling scalable, real-time quantification of Parkinsonian motor symptoms for precision medicine applications.