Explainable Fall Detection for Elderly Care via Temporally Stable SHAP in Skeleton-Based Human Activity Recognition
Explainable fall detection hits 94.3% accuracy under 25ms — but can clinicians actually trust the explanation layer?
The Thesis
A lightweight LSTM + T-SHAP framework delivers real-time skeleton-based fall detection with temporally coherent explanations that clinical staff can act on. Standard SHAP applied frame-by-frame produces jittery attributions that are useless in practice; T-SHAP smooths these over contiguous windows without sacrificing Shapley value guarantees. At 94.3% accuracy on NTU RGB+D with sub-25ms latency on mid-range hardware, this is deployment-credible, not lab-only.
Catalyst
Skeleton-based HAR models have matured enough to run efficiently on edge hardware, and regulatory pressure on AI explainability in clinical settings (EU AI Act, FDA SaMD guidance) has made 'black box accurate' an increasingly untenable product posture. The combination created a real market pull for explainability-first architectures.
What's New
Prior fall detection XAI work either used Grad-CAM (faithfulness AUP: 0.82, weaker than T-SHAP's 0.91) or raw frame-by-frame SHAP (AUP: 0.89), both of which produce temporally unstable attribution maps. T-SHAP applies a linear smoothing operator post-hoc, requiring no retraining and preserving Shapley axioms.
The Counter
The AUP improvement from 0.89 (standard SHAP) to 0.91 (T-SHAP) is a marginal delta — roughly 2 percentage points on a perturbation-based metric that itself has contested validity in the XAI literature. The entire explainability gain might not survive a rigorous clinician user study, which this paper does not include. More fundamentally, skeleton-based detection requires an upstream pose estimation pipeline; in real elder care settings — low-light rooms, partial occlusions, mobility aids, multiple people in frame — pose estimation reliability degrades well before the fall classifier gets a chance to shine. Deploying this in a memory care unit looks nothing like running inference on NTU RGB+D. Finally, the regulatory path for an AI-based fall detection system with clinical decision support claims is long and expensive; the 'strong potential for deployment in clinical monitoring scenarios' language in the abstract does real work to paper over an enormous gap between a benchmark result and a cleared medical device.
Longs
- NVDA
- MSFT
Shorts
- Legacy wearable fall detection vendors (Philips Lifeline, Medical Guardian) whose accelerometer-only products offer zero explainability and lower contextual accuracy
- Grad-CAM-based XAI middleware providers whose benchmark numbers look weaker here
Enablers (Picks & Shovels)
- Azure Kinect / Intel RealSense depth camera vendors for skeleton capture
- NVIDIA Jetson edge inference platform
- AWS HealthLake for clinical data integration
- PyTorch / ONNX runtime for sub-25ms edge deployment
Private Watchlist
- Carepredict
- Vitaltech
- Embodied Intelligence
The Paper
Fall detection in elderly care requires not only accurate classification but also reliable explanations that clinicians can trust. However, existing post-hoc explainability methods, when applied frame-by-frame to sequential data, produce temporally unstable attribution maps that clinicians cannot reliably act upon. To address this issue, we propose a lightweight and explainable framework for skeleton-based fall detection that combines an efficient LSTM model with T-SHAP, a temporally aware post-hoc aggregation strategy that stabilizes SHAP-based feature attributions over contiguous time windows. Unlike standard SHAP, which treats each frame independently, T-SHAP applies a linear smoothing operator to the attribution sequence, reducing high-frequency variance while preserving the theoretical guarantees of Shapley values, including local accuracy and consistency. Experiments on the NTU RGB+D Dataset demonstrate that the proposed framework achieves 94.3% classification accuracy with an end-to-end inference latency below 25 milliseconds, satisfying real-time constraints on mid-range hardware and indicating strong potential for deployment in clinical monitoring scenarios. Quantitative evaluation using perturbation-based faithfulness metrics shows that T-SHAP improves explanation reliability compared to standard SHAP (AUP: 0.89 vs. 0.91) and Grad-CAM (0.82), with consistent improvements observed across five-fold cross-validation, indicating enhanced explanation reliability. The resulting attributions consistently highlight biomechanically relevant motion patterns, including lower-limb instability and changes in spinal alignment, aligning with established clinical observations of fall dynamics and supporting their use as transparent decision aids in long-term care environments