Quantifying Rehabilitation Assessment with Deep Learning

In the rehabilitation of patients with impaired physical function due to conditions such as stroke, the Functional Independence Measure (FIM) is widely used to evaluate independence in activities of daily living. However, because FIM scoring depends on observation and judgment by therapists, it imposes a considerable burden on both patients and healthcare professionals and is subject to inter-rater variability. In this research, we have patients perform simple exercises that differ from the designated FIM assessment actions, and estimate FIM motor item scores automatically from skeletal information extracted from the resulting videos. We use spatio-temporal deep learning models that handle both the spatial relationships among body joints and their temporal changes as a movement unfolds, and incorporate attention mechanisms so that we can identify which body parts and which phases of a movement contribute to the estimate. Through such contribution analysis, we aim to clarify which movements are informative for which assessment items and to build a framework that performs efficient assessment from a small number of simple exercises. Automating part of the assessment would allow therapists to devote more time to interacting with patients, while also providing consistent measures that do not depend on the individual rater. Technical approaches: - FIM motor item score estimation via skeleton estimation and spatio-temporal deep learning - Analysis of contributing joints and movements based on attention mechanisms - Identification of informative movements for each assessment item