Developing Novel Diagnostic and Intervention Schemes for Physical and Neurological Rehabilitation using Deep Learning Methods.

Millions of people suffer from physical and neurological disorders that limit their mobility and

quality of life. In developing countries like Bangladesh, the lack of physiotherapists and advanced

diagnostic tools makes rehabilitation challenging. Traditional methods rely heavily on a doctor’s

experience rather than data-driven approaches. To address this, we developed an AI-powered system

using deep learning to analyze movement patterns and muscle activity for better diagnosis and

rehabilitation planning. Our approach includes two key models: (1) an Auto-Encoder for Motion

Primitive Extraction (AeMP) to decompose complex limb movements into fundamental motion

patterns and (2) an Encoder-Decoder for Muscle Activation Profile Estimation (ED-MAP) to predict

muscle activation from movement data. We conducted experiments using hand and forearm motion

data collected from 23 joint markers and 8 EMG electrodes placed on forearm muscles. Our AeMP

model achieved an average motion reconstruction loss of 0.031 (L1 loss) and successfully extracted

two dominant motion primitives, validated through cosine similarity scores (0.90). These results

indicate the potential of our models for automated rehabilitation assessment. However, our study

is limited by a small dataset and the need for real-world clinical validation. Future work will focus

on expanding the dataset, improving model generalization, and integrating our system into low-cost

rehabilitation devices to enhance accessibility and impact.

• Research Project