Time-varying relevance of functional independence predictors after stroke: a machine learning analysis from the STRATEGY study
Functional recovery after stroke is multifactorial, and predicting medium- to long-term outcomes remains challenging. This study applied machine learning (ML) to data from the prospective observational STRATEGY study to identify, among a minimum set of variables collected at admission to intensive inpatient rehabilitation after stroke, predictors of functional independence at discharge and at 3- and 6-month follow-up. We analysed data from adults admitted within 30 days of stroke to the Florence centre of STRATEGY. Disability in activities of daily living, measured by the modified Barthel Index (mBI) at discharge, at 3 and 6 months after stroke was the outcome of interest. Predictors comprised the minimal assessment protocol for stroke survivors including demographics, premorbid functional scores, neurological and functional variables collected at admission, and a set of markers of clinical complexity in rehabilitation such as dysphagia and urinary incontinence. Three ML algorithms were trained and validated within a nested cross-validation strategy. Performance was evaluated using Median Absolute Error (MedAE). Model interpretability used SHAP values to quantify each predictor’s contribution. Of the 244 recruited patients, 135 had complete outcome data at follow-up. XGBoost yielded the best performance (MedAE 7.83 at discharge, 9.15 at 3 months; 9.52 at 6 months). Baseline mBI was the strongest predictor across all timepoints. Besides mBI, the discharge model attributed greater predictive contributions to acute and subacute factors, including trunk control, neurological severity, lower limb strength, and age and time from the event, whereas 3-month prediction showed greater contributions from premorbid participation, lower limb strength, and cognitive status. At 6 months, neurological severity remained relevant, while premorbid mobility and disability, educational level, and comorbidity burden gained importance. ML models based on a minimal set of relevant clinical and anamnestic variables collected at admission provided accurate predictive performance at all time points. The results support the clinical value of easily and routinely collected variables at admission to inpatient rehabilitation, and show that prognostic factors are phase-dependent rather than static, with clinical and functional measures overarching at discharge, while premorbid disability, global cognitive function, and neurological severity gained importance at longer follow-up. NCT05389878.
Authors
- Chiara Pedrini (ORCID: https://orcid.org/0009-0007-6355-6638)
- Angela Maria Politi (ORCID: https://orcid.org/0000-0002-2801-1436)
- Andrea Mannini (ORCID: https://orcid.org/0000-0003-0760-052X)
- Marco Baccini (ORCID: https://orcid.org/0000-0002-8456-9709)
- Chiara Castagnoli (ORCID: https://orcid.org/0000-0003-1355-2579)
- Francesca Cecchi (ORCID: https://orcid.org/0000-0002-7410-0325)
- Alessandro Sodero (ORCID: https://orcid.org/0000-0003-4356-6912)
- Ilaria Pellegrini (ORCID: https://orcid.org/0000-0002-4736-793X)
- Julieta Giacani
- Ester Marra (ORCID: https://orcid.org/0009-0002-2035-0078)
- Ilenia Ciccone
Institutions
- Don Carlo Gnocchi Foundation (IT)
- University of Florence (IT)
Publication Details
- Journal
- Journal of NeuroEngineering and Rehabilitation
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1186/s12984-026-02216-y
- Primary Topic
- Stroke Rehabilitation and Recovery
- Type
- article
- Field-Weighted Citation Impact
- 0.00