Clinically interpretable prediction models of stroke functional outcomes: A national cohort study of adults in inpatient rehabilitation facilities in the US.

OBJECTIVE

To predict self-care and mobility function at discharge from inpatient rehabilitation for adults with stroke using only variables from the Inpatient Rehabilitation Facility Patient Assessment Instrument (IRF-PAI), which are mandated in the United States by the Centers for Medicare and Medicaid Services.

DESIGN

Retrospective cohort study.

SETTING

Inpatient rehabilitation facilities (IRFs) in the United States.

PARTICIPANTS

National sample (n=43,745) of adults with a primary diagnosis of stroke who were admitted to IRFs in FY2023.

INTERVENTIONS: N/A MAIN OUTCOME MEASURES: Section GG self-care and mobility subscales.

RESULTS

We used random forest regression, an ensemble machine learning approach that trains multiple models and combines their predictions to improve overall performance. We then created global summary trees from the random forest models to visually represent the outcome and aid in clinical interpretation. After data cleaning and quality checks, 39,870 records were available for analysis. Records were divided into training (n=26,580), validation (n=6,645), and test (n=6,645) datasets. Each model included 29 predictors. Random forest models explained 58% (RMSE=5.6) and 62% (RMSE=13.3) of the total variation in self-care and mobility outcomes, respectively, when applied to the test split. In both summary trees, the respective functional measure at admission was the strongest predictor. Bladder and bowel incontinence were strong predictors of both self-care and mobility outcomes at discharge.

CONCLUSIONS

The results show that standard data elements from mandatory Medicare reporting can generate robust, clinically interpretable prediction models of self-care and mobility function at discharge from inpatient rehabilitation for adults with stroke. Application of such models in practice can inform treatment planning and early discharge preparation to support personalized rehabilitation approaches.