A prospective machine learning analysis from the STRATEGY study has identified that the factors predicting functional independence after a stroke are not static, but shift significantly over the course of recovery. Researchers analyzed clinical data from 244 patients admitted to intensive inpatient rehabilitation within 30 days of a stroke, focusing on outcomes at discharge, three months, and six months post-event. By employing a nested cross-validation strategy and training three distinct machine learning models, the team discovered that while baseline functional status is consistently important, the specific variables that best predict long-term independence evolve as patients progress through different phases of recovery.
The study utilized the modified Barthel Index (mBI) to measure disability in activities of daily living across the three follow-up periods. Among the various models tested, the XGBoost algorithm achieved the highest predictive performance, recording a Median Absolute Error (MedAE) of 7.83 at the time of discharge, which slightly increased to 9.15 at three months and 9.52 at six months. This approach allowed the investigators to utilize SHAP values to quantify the contribution of each predictor, transforming a minimal set of routinely collected clinical and anamnestic variables into accurate prognostic tools that remained effective throughout the half-year monitoring window.
At the discharge milestone, the machine learning models indicated that acute and subacute clinical factors held the most predictive weight. Alongside the baseline mBI, variables such as trunk control, neurological severity, lower limb strength, the patient’s age, and the time elapsed since the stroke event were most influential.
As the recovery timeline extended to three months, the model’s predictive focus shifted toward indicators of broader functional and cognitive status. The analysis revealed that premorbid participation, lower limb strength, and the patient’s cognitive status became more prominent drivers of functional outcome predictions.
By the six-month follow-up, the relevance of these predictors transformed again. While neurological severity maintained a level of importance, other variables such as premorbid mobility, pre-existing disability, educational attainment, and overall comorbidity burden emerged as significant contributors to the model’s accuracy.
The researchers emphasize that the predictive power of these models rests on variables that are easily and routinely collected upon admission to inpatient rehabilitation. By demonstrating that prognostic factors are phase-dependent rather than static, the study provides clinical evidence for a more nuanced approach to stroke care.
Despite these promising results, the study acknowledges that of the 244 recruited patients, complete outcome data at the final follow-up was available for 135 individuals. The researchers noted that the findings are derived from the Florence center of the STRATEGY study, which may limit the generalizability of the results to different healthcare systems or populations with vastly different access to rehabilitation resources.
Ultimately, the application of machine learning to the STRATEGY study data reinforces the value of data-driven prognostic modeling in clinical rehabilitation.
Subject of Research: Clinical neuroscience and predictive machine learning analysis in stroke rehabilitation.
Article Title: Time-varying relevance of functional independence predictors after stroke: a machine learning analysis from the STRATEGY study
Article References: Marra, E., Sodero, A., Mannini, A., Baccini, M., Castagnoli, C., Politi, A. M., Giacani, J., Pedrini, C., Pellegrini, I., Ciccone, I., & Cecchi, F. (2026). Time-varying relevance of functional independence predictors after stroke: a machine learning analysis from the STRATEGY study. Journal of NeuroEngineering and Rehabilitation. https://doi.org/10.1186/s12984-026-02216-y
Image Credits: AI Generated
DOI: 10.1186/s12984-026-02216-y
Keywords: stroke recovery, machine learning, functional independence, predictive modeling, STRATEGY study, inpatient rehabilitation, modified Barthel Index, XGBoost, neurological severity, clinical prognosis, long-term outcomes, SHAP values, rehabilitation medicine, post-stroke care, cognitive status
News Source: Teresa Odom. (November 16, 2026). Machine Learning Reveals Dynamic Predictors of Functional Independence Following Stroke. Scienmag.



