Explainable Fall Detection for Elderly Care via Temporally Stable SHAP in Skeleton-Based Human Activity Recognition
In the realm of elderly care, the ability to accurately detect falls is paramount for ensuring the safety and well-being of older adults. However, it is equally important for clinicians to understand the reasoning behind these detections. Traditional methods often fall short in providing reliable explanations of fall events, particularly when utilizing sequential data.
Introduction
This article discusses a novel approach to fall detection that integrates a lightweight framework for skeleton-based human activity recognition with a new explanation strategy known as T-SHAP. This framework aims not only to improve classification accuracy but also to enhance the reliability of explanations provided to clinicians.
The Challenge of Explainability
Existing post-hoc explainability methods, when applied to frame-by-frame analysis of sequential data, often produce temporally unstable attribution maps. This instability makes it difficult for clinicians to trust and act upon the explanations provided. To overcome this challenge, the proposed T-SHAP strategy aggregates feature attributions over contiguous time windows, thereby providing more stable and interpretable insights.
Methodology
The proposed framework combines an efficient Long Short-Term Memory (LSTM) model with T-SHAP, a temporally aware aggregation strategy. Key aspects of this methodology include:
- Linear Smoothing: T-SHAP applies a linear smoothing operator to the attribution sequence, which reduces high-frequency variance in the data.
- Preservation of Guarantees: Despite the smoothing, T-SHAP maintains the theoretical guarantees of Shapley values, such as local accuracy and consistency.
- Real-Time Performance: The framework achieves a classification accuracy of 94.3% with an inference latency of less than 25 milliseconds, making it viable for real-time clinical monitoring.
Experimental Results
Experiments conducted on the NTU RGB+D Dataset reveal that the T-SHAP method enhances explanation reliability compared to standard SHAP and Grad-CAM models. The quantitative evaluations using perturbation-based faithfulness metrics show:
- AUP scores of 0.89 for standard SHAP, 0.91 for T-SHAP, and 0.82 for Grad-CAM.
- Consistent improvements across five-fold cross-validation, reinforcing the reliability of T-SHAP explanations.
Clinical Implications
The resulting attributions from the T-SHAP framework consistently highlight biomechanically relevant motion patterns, such as:
- Lower-limb instability
- Changes in spinal alignment
These findings align with established clinical observations regarding fall dynamics, supporting the framework’s potential as a transparent decision aid in long-term care environments.
Conclusion
The proposed explainable fall detection framework represents a significant advancement in the intersection of AI and elderly care. By providing reliable explanations alongside accurate fall detection, this approach paves the way for improved clinical decision-making and enhanced safety for elderly populations.