Accuracy of Machine Learning to Predict Upper-Limb Outcome Within the First 72 Hours Poststroke

Area of Science:

• Neurology
• Rehabilitation Medicine
• Artificial Intelligence in Medicine

Background:

• Accurate prediction of post-stroke motor outcome is crucial for rehabilitation planning.
• Existing bedside models for upper-limb recovery prediction need improvement for early clinical use within 72 hours of stroke.
• This study focuses on developing a refined prediction model for stroke patients.

Purpose of the Study:

• To develop and internally validate a machine learning model for predicting the 6-month Action Research Arm Test (ARAT) score.
• To utilize simple clinical tests commonly assessed within the first 3 days post-stroke for prediction.
• To enhance the accuracy and feasibility of early post-stroke motor outcome prediction.

Main Methods:

• Utilized data from 296 first-ever ischemic stroke patients across 4 Dutch cohort studies (2000-2019).
• Compared cross-validated prediction performance of multiple eXtreme Gradient Boosting models using various bedside clinical tests.
• Selected a minimal predictor set model balancing feasibility and accuracy, validated on a separate test dataset (n=32) using median absolute error.

Main Results:

• A model incorporating Shoulder Abduction (Motricity Index), voluntary finger extension, Fugl-Meyer Upper Extremity score, and National Institutes of Health Stroke Scale score demonstrated optimal performance.
• The selected model achieved a median absolute error of 5.9 on the 0-57 ARAT score, indicating a good balance between simplicity and predictive accuracy.
• The model's performance was evaluated using median absolute error and interquartile range (2.9-12.9).

Conclusions:

• The developed machine learning model accurately predicts the 6-month ARAT score using a minimal set of bedside clinical tests.
• Predictions are feasible within the first 3 days after stroke.
• The model's median absolute error is below the minimal clinically important difference of 6 points for the ARAT, suggesting clinical relevance for rehabilitation planning.