Group-Based Counterfactual Explanations Aid Rehabilitation Analysis

Interpreting counterfactual explanations (CEs) for multivariate time-series classifiers can be challenging, especially in fields like rehabilitation where experts think in terms of semantic feature groups (e.g., muscle groups) rather than individual sensor channels. Existing counterfactual methods often produce scattered and biomechanically inconsistent explanations when applied to multi-sensor inertial measurement unit (IMU) data used in rehabilitation. This paper proposes a novel two-stage framework to generate group-based counterfactuals for high-dimensional IMU data. It first shows that Shapley-Adaptive (SA) group ranking maintains validity but lacks group-level sparsity. To address this, the researchers introduce Learnable Gate (LG) methods, which incorporate trainable per-group relevance gates optimized alongside perturbation masks. Experiments on the KneE-PAD dataset demonstrate that LG significantly enhances modality-group sparsity and interpretability compared to channel-level baselines, while preserving or improving validity, temporal smoothness, and generation efficiency. This framework provides more actionable, muscle-level corrective guidance for clinicians.