Early Prediction of Discharge Outcomes in Geriatric Rehabilitation: Development and Temporal Validation of a Multiclass Model
Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Objective
To develop and validate a model that predicts on day 7 of subacute rehabilitation (timed to inform the first multidisciplinary team meeting) the probabilities of three mutually exclusive outcomes: minimal dependency, nursing-care dependency, or medical deterioration.
Design
Retrospective cohort study with temporal validation. Multinomial regression and XGBoost models were trained on 80% of admissions during the years 2016-2024 of members of the largest national health maintenance organization (HMO), tested on the remaining 20% and members of other HMOs, and validated on year 2025 admissions. Models output patient-level outcome probabilities with the strongest feature contributors.
Setting
A public geriatric rehabilitation hospital.
Participants
All consecutive first admissions (N=10,920; median age 80 years; median length of stay 35 days).
Interventions
Not applicable.
Main Outcome Measures
Discharge disposition: minimal dependency, nursing-care dependency, or medical deterioration (transfer to acute care without return, or in-hospital death). Predictors were demographic, functional, cognitive, clinical, and laboratory variables documented before day 7 with linkage to pre-admission community health records. Performance was assessed by multiclass area under the receiver operating characteristic curve (AUC), Brier score and positive predictive value (PPV).
Results
Overall, 34% of patients were discharged with minimal dependency, 44% with nursing-care dependency, and 22% deteriorated medically. Both models performed comparably. Multinomial regression achieved an AUC of 0.81 (95% CI: 0.79-0.83) compared with 0.72 (0.71-0.73) for FIM alone, and it was maintained on temporal validation (AUC 0.80, 0.78-0.82). At a ≥60% probability threshold, the model flagged 32% of patients for nursing-care dependency and 5% for deterioration, with PPVs of 72% and 69%, respectively.
Conclusions
Day-7 multiclass risk estimation from routinely collected data is feasible, interpretable and stable over time. Alerts could prompt earlier social-work involvement and closer monitoring of patients at risk of deterioration; prospective controlled evaluation is warranted.
