Computer Vision for Gait Assessment in Cerebral Palsy: Metric Learning and Confidence Estimation.

Area of Science:

• Neurology
• Artificial Intelligence
• Biomedical Engineering

Background:

• Clinical assessment of motor impairments in neurological disorders is crucial but currently relies on time-intensive, qualitative, and clinic-based methods.
• Existing assessments offer limited quantitative detail of patient recovery over time.
• There is a need for accessible, quantitative, and continuous methods for monitoring motor function.

Purpose of the Study:

• To develop and validate a novel machine learning approach for assessing motor impairments using portable devices.
• To improve the accuracy and detail of motor function assessment compared to current clinical methods.
• To enable remote and continuous monitoring of neurological disorder recovery.

Main Methods:

• Leveraged a spatial-temporal graph convolutional network (STGCN) to analyze pose data from monocular video captured by smartphones/tablets.
• Developed an end-to-end model for assessing Cerebral Palsy (CP) using the Gross Motor Function Classification System (GMFCS).
• Incorporated metric learning with triplet loss and self-supervised training for improved performance with limited data and confidence estimation.

Main Results:

• Achieved approximately 76.6% accuracy in assessing children with CP (GMFCS), a 5% improvement over state-of-the-art.
• Demonstrated strong agreement with professional assessments (weighted Cohen's Kappa = 0.733).
• Attained 88% estimation accuracy with a 0.95 confidence threshold, showing effective confidence estimation.

Conclusions:

• The novel STGCN-based approach provides an accurate and efficient method for assessing motor impairments remotely.
• This technology offers quantitative, detailed tracking of recovery, overcoming limitations of traditional clinical assessments.
• The model's efficiency on mobile devices enables real-time or near real-time application in diverse settings, including home environments.