The purpose of this study was to develop and validate a prognostic model to predict discharge FIM, focusing on variables that are observable and compatible with the aging stroke population in Japan.
This study included patients diagnosed with cerebral infarction or hemorrhage who were admitted to the convalescent rehabilitation ward for more than 30 days between January 2018 and December 2020. Exclusion criteria included those with recurrent stroke or worsening condition during hospitalization, a history of neurological diseases other than stroke, and patients with missing data. Variables at admission, such as gender, age, BMI, pre-hospitalization care level, stroke type, lesion location, days from onset to admission, Charlson Index, Food Intake LEVEL Scale (FILS), periventricular hyperintensity (PVH), deep white matter hyperintensity (DWMH), individual FIM item scores (18 items), FIM total score, motor FIM score, and cognitive FIM score, were extracted from past medical records. The primary outcome was the discharge FIM total score, with motor and cognitive FIM scores as secondary outcomes. All data (286 patients) were randomly divided into training data (200 patients) and test data (86 patients). Machine learning was conducted using the random forest algorithm on the training data, utilizing variables collected at admission to predict discharge total FIM, motor FIM, and cognitive FIM scores. Model performance was evaluated using repeated 10-fold cross-validation (K = 10, repeated 3 times), and R², RMSE, and MAE were calculated.
The R², RMSE, and MAE for the prognostic model constructed with the training data were as follows: total FIM (R²: 0.85, RMSE: 11.49, MAE: 8.69), motor FIM (R²: 0.84, RMSE: 8.97, MAE: 6.87), and cognitive FIM (R²: 0.84, RMSE: 3.41, MAE: 2.53). In the test data, the R², RMSE, and MAE were as follows: total FIM (R²: 0.90, RMSE: 10.00, MAE: 7.60), motor FIM (R²: 0.89, RMSE: 8.23, MAE: 6.19), and cognitive FIM (R²: 0.86, RMSE: 3.13, MAE: 2.33).
The prognostic model demonstrated high predictive accuracy, even when using limited, observable variables relevant to an aging population. Further refinement and multi-center validation studies are necessary to improve generalizability and address discrepancies between predicted and actual outcomes.
Despite challenges posed by aging, cognitive decline, and comorbidities, a reliable prognostic model for discharge FIM in stroke patients can be developed, enhancing clinical decision-making in rehabilitation settings.
Functional Independence Measure
Aging stroke population
