Classifying Residual Stroke Severity Using Robotics-Assisted Stroke Rehabilitation: Machine Learning Approach

Area of Science:

• Robotics and Machine Learning in Rehabilitation Science
• Neurorehabilitation Engineering
• Clinical Biomechanics

Background:

• Stroke rehabilitation traditionally occurs in clinical settings, but there's a growing trend towards home-based, technology-integrated recovery.
• This research supports autonomous, in-home stroke recovery by combining robotics and machine learning.

Purpose of the Study:

• To develop supervised machine learning methods for classifying stroke residual severity using in-home kinematics data.
• To improve the accuracy of autonomous classification for stroke rehabilitation.

Main Methods:

• 33 stroke patients used Motus Nova robotics for in-home upper and lower body therapy, collecting motion, assistance, and activity data.
• Data were processed and paired with clinician-defined stroke severity labels (no ROM, low ROM, high ROM).
• Four machine learning algorithms (Light Gradient Boosting, Extra Trees Classifier, Deep Feed-Forward Neural Network, Logistic Regression) were trained and evaluated using an 80:20 data split and 10-fold cross-validation.

Main Results:

• The Light Gradient Boosting (LGB) model demonstrated superior performance with a 96.70% F1-score for autonomous stroke severity detection.
• The LGB model, comprising 139 decision trees, significantly outperformed logistic regression (55.82%), extra trees classifier (94.81%), and deep feed-forward neural network (70.11%).

Conclusions:

• Objective rehabilitation data combined with machine learning can effectively classify residual stroke severity.
• The trained model, utilizing session summary statistics, has potential for real-time integration into clinical settings like outpatient facilities to enhance individualized stroke rehabilitation.