Prediction of Discharge Functional Independence Measure Scores From Routine Admission Variables Using a ChatGPT-Assisted Workflow in Patients With Cerebral Infarction or Intracerebral Hemorrhage: A Retrospective Observational Study.

Background Predicting discharge functional status is important in stroke rehabilitation, but developing local prediction models often requires statistical and programming expertise. The primary objective of this study was to develop and internally validate prediction models for discharge Functional Independence Measure (FIM) scores using routine admission variables. The secondary objective was to describe a locally executed ChatGPT-assisted workflow used only for template generation and code scaffolding. Methods  This single-center retrospective observational study included 377 patients with cerebral infarction or intracerebral hemorrhage admitted to the convalescent rehabilitation ward, a Japanese post-acute inpatient rehabilitation unit for patients who require intensive rehabilitation after acute treatment, of Sakurajyuji Fukuoka Hospital (Fukuoka, Japan) between April 2022 and March 2025. Predictors were age, days from stroke onset to admission, and admission FIM scores. Separate linear regression models were developed for discharge motor FIM and discharge cognitive FIM, and a direct model was developed for discharge total FIM. Internal validation used five-fold cross-validation with out-of-fold predictions. Model performance was evaluated using the mean absolute error (MAE), root mean squared error (RMSE), coefficient of determination (R²), and calibration metrics. Results The discharge motor FIM model yielded an MAE of 11.87, an RMSE of 14.69, and an R² of 0.680. The discharge cognitive FIM model yielded an MAE of 3.76, an RMSE of 4.98, and an R² of 0.731. The direct discharge total FIM model yielded an MAE of 14.55, an RMSE of 18.21, and an R² of 0.714. Calibration slopes were close to 1.0 for all outcomes, and repeated cross-validation showed minimal variation in prediction error. Conclusions Discharge FIM scores were moderately predictable with good calibration using a small set of routinely available admission variables. A locally executed ChatGPT-assisted workflow was feasible for template generation and code scaffolding, but formal evaluation of clinical usefulness, transparent reporting, human verification, and external validation remain necessary.