Digital Signatures of Caregiver Burden: Machine Learning Classification Using Multimodal In-Home Sensor Data
B. Chimehi, J. Larivière-Chartier, B. Wallace, L. Ault, F. Knoefel, J. Kaye, Z. Beattie, J. Steele, L. Anderson, N.Thomas. Gerontechnology 25(s)
Full text PDF 
( Download count: 1)
AbstractPURPOSE: Caregiver burden is typically measured through self-report, and has not been studied using objective, continuous indicators of daily in-home activity. Sensor-based monitoring platforms, such as those developed through the Oregon Center for Aging & Technology (ORCATECH) [1], offer an opportunity to unobtrusively capture daily patterns and caregiving-related behaviors. While feasible, it remains unclear which sensor modalities best correspond to caregiver burden in dyads affected by cognitive impairment. This study aims to determine the sensor-based outcomes that are associated with caregiver burden levels using machine learning models trained on multimodal sensor combinations collected over 18 months. METHODS: 47 spousal dyads were enrolled, each consisting of a person with cognitive impairment (PWCI) and their care partner. Following baseline assessments, the ORCATECH platform was installed in each home. Motion sensors, exit door contact sensors, bed sensors, wearable devices, a weight scale, and medication trackers provided continuous, high-frequency data on mobility, sleep, and health-related behaviors. Care partners completed weekly Zarit Burden Interview Short Form (ZBI-12) surveys [2]. RESULTS AND DISCUSSION: A total of 1,122 multisensor feature combinations were generated and evaluated based on classification accuracy and data completeness. 44 dyads contributed valid weekly data. The top-performing ML model was a Decision Tree Classifier trained on features from motion sensor and smartwatch steps from PWCI and their caregiver. This model achieved a weighted accuracy of 72%, outperforming all single-sensor and alternative configurations. Sensors measuring daily activity levels and environmental motion data produced the strongest predictive capability, supporting the potential of passive in-home sensing to identify behavioral markers of caregiver burden. These findings suggest that passive in-home sensing can enable early identification of elevated caregiver burden and facilitate timely implementation of support services before caregivers become overwhelmed.Keywords: Person with cognitive impairment (PWCI), Dementia, Caregiver burden, Sensor, ZBI.
B. Chimehi, J. Larivière-Chartier, B. Wallace, L. Ault, F. Knoefel, J. Kaye, Z. Beattie, J. Steele, L. Anderson, N.Thomas. Gerontechnology 25(s) (2026). Digital Signatures of Caregiver Burden: Machine Learning Classification Using Multimodal In-Home Sensor Data. Gerontechnology, 25(2), 1-10
https://doi.org/10.4017/gt.2026.25.2.1465.3