An Interpretable AI for Smart Homes: Identifying Fall Prevention Strategies for Older Adults Using Multimodal Deep Learning.

Abstract

Falls are a significant cause of mortality among older adults and are considered preventable emergencies. We developed an interpretation framework, including multimodal predictive models that capture both static and time-series data, and analyzed global feature importance via a perturbation approach, permutation importance (PIMP), and SHapley Additive exPlanations (SHAP) to identify fall prevention strategies at home. Our predictive model utilizing the BiCrossNet architecture achieved a prediction accuracy of 98% when 12,540 data points were used. It was found that activity-related and indoor thermal environment features are important for predicting fall emergency occurrences (FEOs). Interestingly, extended active time in the living room/kitchen and outdoors decreases the FEO ratio, whereas extended active time in the bedroom and bathroom increases the FEO ratio. For example, increasing the active time in the living room/kitchen from 60-70 to 130-140 minutes reduced the FEO ratio by 32% . Furthermore, indoor temperature and humidity were found to be important depending on the outdoor conditions. Moreover, this study shows the importance of personalized prevention strategies depending on the underlying health conditions of older adults.