XAI for Transformer-based Fall Detection using Skeleton Time-series Data Analysis

Accurate fall detection is increasingly important for continuous monitoring in home and workplace environments. Recent advances have focused on vision-based systems and skeleton-based representations, exploiting Transformer-based architectures to model complex spatio-temporal relationships. However, the lack of transparency of these models limits trust and practical deployment. To address this challenge, this paper employs ShaTS, a variant of the SHAP technique, as a post-hoc, model-agnostic explainability method to interpret and understand the predictions of the Transformer-based model. ShaTS is used to quantify the global contributions of joints and temporal segments on the model's predictions. The results show that explainability can reveal meaningful motion cues, support domain-level validation, and guide future model refinement, ultimately enhancing confidence in the model's decision-making process.