Can machine learning improve on the early prediction of upper limb recovery after stroke?

BACKGROUND: Early prediction of upper limb recovery is important to optimise rehabilitation and inform patients but remains challenging due to inter-individual variability. This study aims to (1) develop and validate a machine learning model to predict arm-hand capacity at six months post-stroke using clinical variables from the first week; (2) compare its performance to a mixed-effects model; and (3) co-design a user-friendly output visualisation with clinician input. METHODS: From data of 451 first-ever ischemic stroke patients, we selected total Action Research Arm Test score (ARAT), shoulder abduction, and finger extension as predictors. An XGBoost model was trained on these variables measured at varying time points within the first five months, using 5-fold, 5-repeat cross-validation. We employed bootstrap aggregation to obtain generalisable predictions and prediction intervals to quantify uncertainty. The model's performance was validated on a hold-out set and compared against a mixed-effects model using median absolute error (MedAE). RESULTS:  The XGBoost model achieved a MedAE of 4.2 points (IQR = [1.2, 12.6]) on the ARAT when applied at seven days post-stroke, compared to 13.7 points (IQR = [4.6, 27.8]) for the mixed-effects model in the same patients. CONCLUSION: Our model provides significantly more accurate predictions of upper limb recovery, with a 69% error reduction compared to the mixed-effects model. Its ease of use, interpretability, and use of routinely collected clinical data make it suitable for digital clinical workflows. Future research could validate the model in larger, more recent cohorts and explore integrating neuroimaging and temporal features.