Understanding the Importance of Feature Groups for Clinical Outcome Predictions with Machine Learning in Post-Stroke Robotic-Assisted Rehabilitation

Abstract

Outcome predictions in post-stroke rehabilitation are a key element to personalize the treatment to the needs of the patient, finally enhancing effectiveness of the therapy. They can form the basis of Decision Support Systems, helping clinicians to progressively tune the therapy depending on patients' clinical status and progress. Diverse data sources, such as clinical, demographic, kinematic and time-related data in robotic-assisted rehabilitation, can provide different prediction results. Understanding which data source, or combination thereof, contains useful information for outcome predictions can improve the development of machine learning tools, Decision Support Systems, and even clinical setups designed to record these useful data. The presented work investigates different feature groups and machine learning methods, using data recorded within a robotic-assisted rehabilitation treatment including 44 stroke patients. Results highlight the effectiveness of using multi-dimensional feature groups to predict poststroke rehabilitation. While clinical data alone can already achieve a solid basis for predictive modeling, the integration of kinematic and time-related data can significantly improve prediction accuracy of the patient outcome.