Activity routine discovery in stroke rehabilitation patients without data annotation

Abstract

In this work, we investigated whether activity routines of stroke rehabilitation patients can be discovered from body-worn motion sensor data and without data annotation using topic modeling. Information about the activity routines performed by stroke patients during daily life could add valuable information to personal therapy goals. As topic model input, we used a set of activity primitives derived from upper and lower extremity motion sensor data. We monitored three stroke patients during their daily life in a day care center for 8 days each within 3 weeks. We achieved up to 88% accuracy for activity routine discovery for subject-dependent evaluations. Our discovery approach seems suitable for activity routine discovery in rehabilitation patients.