Prediction of safety accident subtypes for persons with dementia using sensors and machine learning: an observational study

Abstract

Background and Objectives We explored the use of machine learning models for predicting safety accident subtypes among individuals with dementia using in-home sensors and to identify key predictors.Research Design and Methods An observational study was conducted using 966 days of in-home sensor data, sleep data from wearable Actiwatch devices, caregiver-completed structured safety accident diary data, and individual data collected in South Korea. Five machine learning classification models were developed to predict physical injury, nighttime behaviors/wandering, and risky behaviors. Model performance was compared, and the most important predictive features were extracted.Results The Gradient Boosting Machine showed the best performance in predicting physical injury and nighttime behaviors, while CatBoost performed best for risky behaviors. Activity patterns recorded using in-home sensors emerged as essential features for predicting different safety accident subgroups, particularly for nighttime behaviors and wandering.Discussion and Implications These findings highlight the potential of these technologies to identify high-risk individuals with dementia. Further research is recommended to integrate these methods for daily safety monitoring of this population.