Machine Learning Risk Prediction for Prolonged Hospitalization in Frail Older Adults with Multimorbidity
Discuss this preprint
Start a discussion What are Sciety discussions?Listed in
This article is not in any list yet, why not save it to one of your lists.Abstract
Background
Frailty and multimorbidity are common in older adults and contribute substantially to prolonged hospitalizations, readmissions, and mortality. Yet, existing prediction models often fail to integrate frailty-specific biomarkers and lack interpretability for routine clinical use.
Objectives
To develop and internally validate an interpretable, machine learning–enhanced logistic regression model to predict prolonged hospital length of stay (LOS) among frail older adults with multimorbidity, and to identify key predictors to guide individualized inpatient care.
Methods
We conducted a retrospective study of 440 hospitalized adults aged ≥65 years with multimorbidity (≥2 chronic conditions) and frailty (Frailty Index ≥0.25) at a tertiary geriatric department between January 2022 and December 2023. Fourteen demographic, clinical, and biochemical variables were analysed. Feature selection employed Elastic Net regularization, Extreme Gradient Boosting with SHAP value analysis, and the Boruta algorithm to ensure robust predictor identification. A multivariable logistic regression model was trained and internally validated using stratified 10-fold cross-validation and 1,000 bootstrap iterations. Discrimination (AUC-ROC), calibration, and clinical utility (decision curve analysis) were assessed.
Results
Eight predictors age, diabetes, hypertension, prior stroke, serum albumin, HDL cholesterol, systolic blood pressure, and neutrophil-to-lymphocyte ratio—were retained in the final model. The model achieved good discrimination (AUC = 0.770, 95% CI 0.688–0.853) and acceptable calibration (Hosmer–Lemeshow χ² = 14.86, p = 0.062). Cross-validation (mean AUC 0.687 ± 0.072) and bootstrap correction (AUC 0.672) confirmed internal stability. Serum albumin was the strongest protective factor, while elevated neutrophil-to-lymphocyte ratio and prior stroke were significant risk factors.
Conclusions
This interpretable model accurately predicts prolonged hospital stay in frail older adults with multimorbidity using routinely available clinical data. Its transparent design supports integration into electronic health records for real-time risk stratification, facilitating targeted discharge planning and personalized geriatric care.
