Machine-learning prediction of 6-month unplanned hospitalization among long-term care residents: A multicenter study...

Unplanned hospitalizations are frequent and costly among residents of long-term care facilities, but prediction tools based on routine facility data remain limited. We developed a machine-learning model for 6-month unplanned hospitalization and evaluated its performance through institutional external validation across institutionally distinct LTCF cohorts within one province.

We conducted a multicenter retrospective cohort study in 15 LTCFs in Anhui, China, between March 2024 and April 2025. Prespecified variables from routine assessments were considered, and feature selection yielded 11 retained predictors. Five machine-learning algorithms were developed and evaluated in a development set, a held-out internal test set, and two external validation cohorts. Model performance was assessed by discrimination, calibration, decision curve analysis, and model interpretation.

Among 768 residents, 253 (32.9%) experienced a first unplanned hospitalization within 6 months. Random forest showed the best performance, with areas under the curve of 0.932 in the development set, 0.934 in the held-out internal test set, 0.953 in external validation set 1, and 0.869 in external validation set 2. Calibration varied across cohorts, with slope-level miscalibration in external validation set 1 and underprediction in external validation set 2. Decision curve analyses indicated positive net benefit across threshold probabilities of 5%-30% in the held-out internal test set and both external validation cohorts.

The model showed useful discriminative ability but variable calibration across LTCF settings. It may support future risk stratification using routinely collected data, but broader geographic validation and prospective implementation studies are needed.