Hand weakness is a leading cause of long-term disability among stroke survivors, severely limiting daily activities and quality of life. While wrist-worn accelerometers offer an objective means of measuring upper limb use in real-world settings, traditional metrics — such as movement duration and interlimb ratios — provide limited clinical insight. When combined with unsupervised clustering, these heuristic measures frequently produce groups with substantial overlap on validated clinical scales like the Action Research Arm Test (ARAT).
This paper presents a supervised machine learning framework that classifies post-stroke upper limb performance from 24-hour accelerometer recordings in a way that aligns with clinically validated ARAT scores. A deep neural network (DNN) extracts latent features directly from raw triaxial accelerometer data, which are then used alongside demographic and heuristic accelerometry variables in a random forest classifier to categorize participants into five distinct performance groups.
The study analyzed data from 211 participants across three cohorts: individuals with acute/subacute stroke (n = 57, tracked 2 weeks to 6 months post-stroke), individuals with chronic stroke (n = 78, more than 6 months post-stroke), and neurologically intact adults serving as controls (n = 76). Participants wore ActiGraph GTX3 accelerometers on both wrists during 24-hour periods outside clinical settings.
Participants were grouped into five ARAT-based functional categories: No Function (0–10), Poor (11–21), Limited (22–42), Notable (43–54), and Full (55–57). Per-second acceleration magnitudes were computed and segmented into 30-minute windows, with segments showing over 90% inactivity removed. A DNN with hidden layers of sizes 128, 32, and 8 learns an 8-dimensional latent representation per segment; participant-level representations are formed by averaging across all valid segments. These latent features, combined with five heuristic accelerometer variables (paretic limb duration, non-paretic limb duration, median paretic limb acceleration, paretic limb variability, and use ratio) and demographic variables, are fed into a random forest classifier evaluated via 10-fold cross-validation.
Traditional stroke recovery assessment relies on periodic clinical visits using standardized tasks in controlled environments. These episodic snapshots may not reflect how patients actually use their affected hand in daily life, and gains measured in the clinic frequently do not transfer to real-world function. This work bridges that gap by aligning continuous, passive wearable sensor data with validated clinical capacity scores, enabling each ARAT score to map to a distinct real-world performance profile.
The framework is designed for scalability and automation — reducing reliance on manual clinical assessments while providing patients and clinicians with actionable, data-driven insights to support personalized rehabilitation planning and outcome monitoring. Future directions include incorporating recurrent or transformer architectures to better capture sequential dynamics, and external validation on independent datasets to support broader clinical adoption.
This research was conducted in collaboration with the Department of Occupational & Recreational Therapies, the Kahlert School of Computing, the Department of Mechanical Engineering, and the Department of Biomedical Engineering at the University of Utah, as well as the Department of Electrical & Computer Engineering at the University of Pittsburgh, under the supervision of Dr. Amit Sethi.