Explaining the unique nature of individual gait patterns with deep learning

Area of Science:

• Biomechanical analysis
• Machine learning applications
• Clinical diagnostics

Background:

• Machine learning (ML), including deep neural networks (DNNs), excels at complex system modeling but often acts as a "black box."
• This lack of transparency hinders clinical acceptance, particularly in medical diagnoses.
• Understanding the predictive basis of ML models is crucial for reliable application in healthcare.

Purpose of the Study:

• To investigate the uniqueness of individual gait patterns in clinical biomechanics using DNNs.
• To develop a method for interpreting the predictions of ML models in gait analysis.
• To enhance the clinical applicability of ML in diagnosing and treating human gait disorders.

Main Methods:

• Application of deep neural networks (DNNs) to analyze gait data.
• Utilizing Layer-Wise Relevance Propagation (LRP) to attribute model predictions to input variables.
• Analysis of ground reaction forces and full-body joint angles during the gait cycle.

Main Results:

• Layer-Wise Relevance Propagation (LRP) successfully identified relevant input variables and time windows for characterizing individual gait patterns.
• The study demonstrated which biomechanical variables are most critical for distinguishing unique gaits.
• A general framework for interpreting non-linear ML methods in biomechanical gait analysis was established.

Conclusions:

• The developed method provides a transparent approach to understanding ML predictions in gait analysis.
• This interpretability is vital for the clinical acceptance and application of ML in diagnosing and treating human gait.
• The framework offers a powerful tool for advancing clinical biomechanics and patient care.