Smartphone-Based Brunnstrom Stage Classification of Hemiparetic Gait via Skeleton-Attention-LSTM-Inception Network

Area of Science:

• Biomechanics and Rehabilitation Engineering
• Artificial Intelligence in Healthcare
• Neuroscience and Motor Control

Background:

• Accurate assessment of hemiparetic gait post-stroke is crucial for understanding motor impairment and guiding rehabilitation.
• Current Brunnstrom Recovery Stage (BRS) assessments can suffer from inter-rater variability, impacting longitudinal tracking.
• Quantitative methods are needed to objectively monitor stroke recovery and personalize therapy.

Purpose of the Study:

• To develop and validate a deep learning classifier for objective, quantitative assessment of hemiparetic gait.
• To eliminate inter-rater variability in Brunnstrom Recovery Stage (BRS) classification.
• To enable precise longitudinal tracking of motor recovery after stroke.

Main Methods:

• Collected kinematic data from 40 healthy adults and 51 stroke patients during walking tasks.
• Utilized advanced pose estimation to extract 3D keypoint coordinates for fine-grained kinematic analysis.
• Developed a hybrid Skeleton-Attention Long Short-Term Memory (LSTM)-Inception architecture for BRS stage classification, integrating temporal and spatial feature extraction.

Main Results:

• The proposed deep learning framework achieved a superior classification accuracy of 97.3%.
• This significantly outperformed conventional deep learning benchmarks (87.7-95.2%).
• Demonstrated high reliability and accuracy in classifying BRS stages from kinematic data.

Conclusions:

• Deep learning-driven motion analytics can effectively eliminate subjectivity in Brunnstrom Recovery Stage (BRS) staging.
• This approach facilitates quantitative longitudinal monitoring of neurorehabilitation progress.
• Enables data-driven personalized therapeutic strategies, potentially reducing reliance on clinician expertise.