Classification Models of Action Research Arm Test Activities in Post-Stroke Patients Based on Human Hand Motion.

Area of Science:

• Biomedical Engineering
• Rehabilitation Science
• Machine Learning

Background:

• The Action Research Arm Test (ARAT) exhibits a ceiling effect, limiting its sensitivity in detecting functional improvements in stroke survivors with mild hand impairments.
• Accurate assessment of fine motor skills is crucial for effective stroke rehabilitation.

Purpose of the Study:

• To develop and evaluate machine learning models for differentiating between healthy individuals and post-stroke subjects based on finger joint kinematics.
• To address class imbalance in datasets for improved classification performance in stroke research.

Main Methods:

• Utilized extension and flexion angles of 11 finger joints as features for classification.
• Employed Support Vector Machine (SVM), Random Forest (RF), and K-Nearest Neighbors (KNN) algorithms.
• Implemented Borderline-SMOTE for addressing class imbalance.

Main Results:

• Data balancing with Borderline-SMOTE significantly enhanced classification metrics (accuracy, recall, f1-score, AUC) for all models.
• The SVM classifier achieved superior performance post-balancing, with 98% precision, 97.5% recall, and 0.996 AUC.
• Classification models leveraging hand motion features and Borderline-SMOTE demonstrated high efficacy.

Conclusions:

• Machine learning models incorporating finger joint kinematics and oversampling techniques can effectively identify subtle functional differences in post-stroke individuals.
• These advanced methods offer a more sensitive approach to assessing hand function recovery compared to traditional clinical scales like the ARAT.