Clinically Interpretable XAI-Based FMCW Radar Fall Detection with Reduced False Alarms

AI-based fall detection systems often limit clinical trust due to high false alarm rates and opaque decision-making. This study presents a novel Explainable AI (XAI) framework designed for FMCW radar-based fall detection. Guided by Patient and Public Involvement, a dataset of 400 falls and 500 non-fall daily activities was collected from volunteers aged 26 to 78 in controlled environments and private homes. Multiple machine learning models, including Random Forest, SVM, XGBoost, and Neural Networks, were benchmarked before and after integrating the proposed XAI framework using 35 micro-Doppler (MD) extracted features. Initially, biomechanically relevant radar features contributing to classification decisions were identified using Permutation Importance. Then, a Surrogate Decision Tree was trained to approximate the optimized model, generating human-readable classification rules for clinicians and caregivers. Finally, a counterfactual analysis module was designed to identify feature shifts in false positives and false negatives. These shifts and surrogate rules guided threshold refinements to improve decision boundaries, reducing false alarms and missed falls. Results show that false alarms were often associated with excessive MD width, while low-energy falls contributed to missed detections. Integrating the XAI framework led to improvements across all classifiers, with model accuracy increasing by 4.9% to 9.6%, false alarm rates decreasing by 1.0% to 9.0%, and missed fall rates reducing by 1.3% to 8.8%. As proof of concept, the framework has been clinically validated with expert feedback, demonstrating strong potential for real-world deployment.