To develop the Primitive Sequencing pipeline (PrimSeq), a deep learning-based framework to automatically identify and count functional primitives in rehabilitation training, investigators captured the upper body motion of 41 chronic stroke patients aged 21 to 84 years old who performed a battery of rehabilitation activities. Eligibility criteria included premorbid right-handedness, ability to give informed consent, and unilateral motor stroke with contralateral upper extremity weakness scoring <5/5 in any major muscle group. Patients were excluded if they had hemorrhagic stroke with mass effect, or subarachnoid or intraventricular hemorrhage; traumatic brain injury; a musculoskeletal, major medical, or non-stroke neurological condition that interferes with motor function; contracture at shoulder, elbow, or wrist; moderate UE dysmetria or truncal ataxia; apraxia; visuospatial neglect; global inattention; or legal blindness. Upper extremity impairment level was measured using the Fugl-Meyer Assessment (UE-FMA). The functional primitives explored in this project were reach, reposition, transport, stabilize, and idle motions.

The dataset includes biomechanical data captured by 9 inertial measurement units (IMU) affixed to each stroke patient's upper body, video capture from 2 orthogonal cameras, patient demographic and clinical information, as well as IMU data from healthy controls.