Leveraging machine learning for digital gait analysis in ataxia using sensor-free motion capture.

Area of Science:

• Neurology
• Biomedical Engineering
• Data Science

Background:

• Gait disturbances are a hallmark of ataxia, with current clinical scales offering limited precision in assessing severity and individual deterioration.
• Sensor-free motion capture technology presents a novel approach to quantitatively analyze gait, potentially overcoming the limitations of traditional clinical assessments.

Purpose of the Study:

• To evaluate the efficacy of sensor-free motion capture combined with machine learning (ML) in replicating and enhancing the assessment of gait disturbances in ataxia patients.
• To determine if ML models can detect subtle and longitudinal changes in gait that are not apparent with current clinical scoring methods.

Main Methods:

• Utilized AlphaPose for full-body pose estimation from videotaped walking tasks of 91 ataxia patients and 28 healthy controls.
• Employed machine learning models (tsfresh, ROCKET, XGBoost, Ridge) to analyze time-series data from motion capture for gait assessment.
• Applied explainable AI (SHAP) to identify key gait parameters influencing ML model predictions.

Main Results:

• ML models achieved high accuracy in gait disturbance assessment, outperforming human clinical ratings in categorical prediction (F1-score 63.99% vs. 60.57%).
• Successfully differentiated pre-symptomatic ataxia patients from healthy controls (F1-score 75.96%), highlighting sensitivity to subtle changes.
• Demonstrated significant accuracy in detecting longitudinal gait changes over time (Pearson's r = -0.626, p < 0.01), unlike human assessment (r = -0.060).

Conclusions:

• ML-based analysis of sensor-free motion capture provides a more sensitive method for assessing gait disturbances in ataxia.
• This approach can capture subtle and longitudinal gait alterations, offering potential for early intervention and therapy monitoring.
• Findings suggest promise for these methods as outcome measures in clinical trials and for home-based patient monitoring.