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

Timely and accurate prediction of poststroke motor outcome is important for efficient rehabilitation planning and resource allocation. Existing bedside models for predicting upper-limb outcome after stroke require further refinement to be effectively implemented in stroke units within the first 72 hours. This study aimed to develop and internally validate a machine learning model to predict the 6-month Action Research Arm Test score using simple clinical tests commonly assessed within the first 3 days poststroke.

In 296 first-ever ischemic stroke patients pooled from 4 prospective Dutch cohort studies across 44 centers (2000-2019), we compared the cross-validated prediction performance of multiple eXtreme Gradient Boosting models using different sets of bedside clinical tests to predict the 6-month Action Research Arm Test outcome (0-57). We then selected the model with the minimal predictor set that best balanced bedside feasibility and accuracy and validated it within 72 hours poststroke on a test data set (n=32) from the same cohort using median absolute error as the evaluation metric.

A model incorporating Shoulder Abduction from the Motricity Index, voluntary finger extension, Fugl-Meyer Upper Extremity, and total National Institutes of Health Stroke Scale score as bedside tests, showed the best tradeoff between model simplicity and predictive accuracy (median absolute error, 5.9 on the 0-57 score; interquartile range, 2.9-12.9).

Our model predicts the 6-month Action Research Arm Test score using a minimal set of bedside clinical tests collected within the first 3 days after stroke, achieving a median absolute error below the Action Research Arm Test minimal clinically important difference of 6 points.