From Motion Data to Meaning: A Dual-User Explainable AI Framework for Human-Centered Biomechanical Analysis
Movement analysis is a critical component of clinical biomechanics, rehabilitation, and motor control research. Although advances in markerless motion capture have enabled the extraction of highly detailed biomechanical data, this information remains largely underutilized in standard clinical practice. The primary barrier is an "interpretation gap": biomechanical signals are inherently high-dimensional and abstract. While machine learning (ML) models excel at pattern recognition in these complex datasets, their "black-box" outputs lack the transparency necessary for clinical trust and decision-making. To address this, we present an end-to-end computational framework and dual-user interface designed to transform complex motion data into actionable, role-specific insights. ; ; Our framework utilizes a Two-Stream Network that serves as the backend for the interface system, processing data from clinical assessment tasks (e.g. Gait). It employs two complementary clustering approaches: ; ; - Discrete Clustering: Groups subjects based on demographic data (age, sex, weight) and performance features like range of motion (ROM) and joint strength. ; - Time-Series Clustering: Applies Dynamic Time Warping (DTW) to normalized kinematic curves (101-frame trajectories) to identify execution strategies. ; ; Because traditional gradient-based XAI methods (like standard SHAP or Grad-CAM) assume differentiable models and are incompatible with distance-based DTW clustering, we implemented two adapted techniques: ; ; - Perturbation-based Feature Attribution (Pseudo-SHAP): Calculates feature importance by measuring normalized deviations from cluster baselines, attributing specific risk factors to outcomes without game-theoretic estimation. ; - Prototype-based Attention Maps: Computes frame-wise and feature-wise Euclidean distances from the cluster prototype to highlight the specific temporal phases driving the classification. ; ; Rather than treating interpretability as an afterthought, XAI logic is integrated directly into the real-time classification engine. The system transforms algorithmic outputs into role-specific representations for two distinct user personas: ; ; - The Clinician View: Designed for analytical triage and investigation, this module provides risk decomposition and kinematic attention heatmaps that overlay specific gait phases triggering alerts. ; - The Patient View: To foster engagement, complex XAI matrices are simplified via counterfactual analysis into actionable visual feedback, including "Target Zone" charts and a 0–100 progress score to build a narrative of recovery. ; ; We will evaluate this system through a Human-in-the-Loop validation methodology to capture interaction dynamics. By combining the System Usability Scale (SUS), observational analysis of workflow disruption, and qualitative trust assessments, this framework aims to ensure that visual explanations align with expert intuition while preserving human agency.
Back to Current Students