Deep Learning-based Data-Efficient Stroke Gait Analysis: A Focused Review

Stroke is a leading cause of disability worldwide, often causing gait impairments that limit a person’s mobility and independence. Accurate gait analysis is critical for effective rehabilitation, but traditional evaluation approaches are based on clinical observations that lack precision and consistency and are affected by the physiotherapist’s experience. Deep learning models, especially Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM) networks, have been shown to have potential in gait analysis but their reliance on huge datasets restricts its relevance to stroke rehabilitation since patient-specific gait patterns have significant variability. This review explores the potential of current state-of-the-art models utilizing motion sensor data (IMU) towards learning individual strokesurvivors’ unique gait characteristics, to be studied by physiotherapists through simulation for developing individualized recovery plans. This study recommends CNNLSTM architecture, Few-Shot Learning (FSL) approaches, and synthetic data generation to address data scarcity and enhance model personalization for stroke gait analysis. However, challenges such as model complexity, clinical interpretability, lack of resolution in IMU sources and the reliability of synthetic data are some of the highlighted obstacles that demand attention for future work.