Explainable and Uncertainty-Calibrated Decision Making for Fall Detection with Edge Deployment
Students & Supervisors
Student Authors
Supervisors
Abstract
Falls cause severe health issues among the aging population, requiring viable, real-time fall detection systems to be deployed in clinical settings. This study introduces an explainable, uncertainty-calibrated fall detector deployed at the edge, addressing the major shortcomings of existing methodologies in terms of interpretability, uncertainty quantification, and computational concerns. The proposed framework combines evidential deep learning with lightweight backbones (ResNet18/ MobileNetV3) and temporal attention to process 16 -frame sequences and make calibrated uncertainty estimates using the Dirichlet distribution. The system achieved a test accuracy of 96.73% on the CAUCA-Fall dataset and 95.65% on the Le2i dataset, which is clinically acceptable and demonstrates robust generalization between controlled and realistic settings. Critical fall detection has a precision of 98 % and a recall of 97 % for difficult real-world examples. The excellent values of the Expected Calibration Error (ECE <0.03) with uncertainty calibration allow the adaptive use of thresholding and multi-level alert systems. Decision auditability is explained by visualizations of GradCAM and temporal attention weights, which are interpretable by clinicians. The system has shown considerable computational gains over the baselines, with uncertainty quantification and explainability features required for deployment in healthcare settings, helping to bridge the significant research-to-clinicalapplication gap.
Keywords
Publication Details
- DOI: 10.1109/ICCIT68739.2025.11491588
- Type of Publication:
- Conference Name: 28th International Conference on Computer and Information Technology (ICCIT)
- Date of Conference: 19/12/2025 - 19/12/2025
- Venue: Cox's Bazar, Bangladesh
- Organizer: IEEE