Predicting Caregiver Burden Using Motion Sensor Data and Machine Learning: A Supportive Approach for Persons with Cognitive Decline

Alzheimer's disease (AD) is a progressive neurodegenerative disorder that impacts individuals and their caregivers. As AD advances, caregivers experience heightened levels of burden, often exacerbated by changes in the behavioral patterns of persons with cognitive impairment (PWCI). In this study, we leverage in-home motion sensor data to analyze activity patterns and predict caregiver burden using machine learning. The dataset is captured with the Oregon Center for Aging & Technology sensor platform, which used ambient sensors to continuously monitor movements in the home and specifically room usage for participants over 18 months. We evaluated 5 machine learning models using various preprocessing options and found that Decision Tree Classifier using Independent Component Analysis (ICA) preprocessing for feature transformation the most effective for burden classification. The ICA extracted three independent components that highlight distinct behavioral patterns. To further interpret the model's predictions, we utilized SHAP (SHapley Additive exPlanations) values, which quantify the impact of individual features on the classification outcome. The analysis revealed key behavioral indicators, such as time spent in specific rooms and frequency of room visits correlate with higher caregiver burden levels. Our results show that more time spent in common areas, such as the living room, is associated with lower caregiver burden, while increased visits to the bathroom are linked to higher burden levels. The model achieved an accuracy of 70% with ICA-transformed features, demonstrating the potential of sensor-driven analytics in predicting caregiver burden. These findings underscore the value of machine learning-based monitoring in providing objective, data-driven assessments to support caregivers and improve dementia care strategies.