Enhancing stroke recovery assessment: A machine learning approach to real-world hand function analysis

Hand weakness is a major contributor to long-term disability in stroke survivors, severely affecting daily function and quality of life. Although wrist-worn accelerometers offer an objective means of measuring upper limb (UL) use in daily life, traditional metrics such as movement duration and interlimb ratios provide only limited insight. When combined with unsupervised clustering, these heuristic measures often fail to capture meaningful clinical differences as the groupings frequently show substantial overlap on clinical scales like the Action Research Arm Test (ARAT).

To develop a machine learning-based method for categorizing post-stroke upper limb performance from 24-hour accelerometer recordings in a way that aligns with clinically validated ARAT scores.

This study analyzed continuous 24-hour triaxial accelerometer data to capture real-world upper limb movement. A deep neural network was applied to extract features directly from the raw data. These learned features were then used to categorize participants into five performance groups based on ARAT scores, aiming to better reflect clinically meaningful functional recovery.

The proposed method produced non-overlapping groupings of upper limb performance aligned with ARAT-defined categories. It achieved a classification accuracy of 97%, significantly outperforming models based on demographics and heuristic accelerometry features, which reached only 66%.

This study presents a machine learning framework that accurately classifies clinically relevant real-world upper limb function using accelerometer data. This approach offers a more precise and objective assessment of post-stroke hand use, with potential to support personalized rehabilitation planning and outcome monitoring.