Ensemble Machine Learning Models for Predicting Death in Post-Acute Skilled Nursing Facilities.

OBJECTIVE: Patients discharged from hospitals to skilled nursing facilities (SNFs) for post-acute care are at high risk for adverse outcomes, including death. We aimed to develop machine learning (ML) models that predict death within 30 or 180 d after patients are discharged to SNFs. DESIGN: Retrospective cohort study. SETTING AND PARTICIPANTS: Patients discharged to any of 75 SNFs in our region between October 1, 2019, and June 30, 2021, after being hospitalized at the Mayo Clinic in Rochester, Minnesota, or Mayo Clinic Health System. METHODS: Ensemble ML models for predicting 30- or 180-d death were developed and trained by using supervised learning. Each model comprised 5 independent classifiers: random forest, extreme gradient boosting, category boosting, extra trees, and light gradient-boosting machine. Outputs of the classifiers were then aggregated and passed to a voting mechanism. Data were split so that 70% were used for training, 10% for hyperparameter tuning, and 20% for testing in a 10-fold cross-validation scheme. Model performance was assessed by using area under the curve (AUC), and 95% CIs were calculated with bootstrap sampling. Contributions of individual variables were assessed with Shapley additive explanations. RESULTS: We identified 7103 SNF admissions of 5510 individual patients. After admission to an SNF, 337 patients (4.7%) died within 30 d, and 1328 patients (18.7%) died within 180 d. The models performed well, with AUCs of 0.84 (95% CI, 0.74-0.92) and 0.82 (95% CI, 0.73-0.90) for predicting death within 30 or 180 d, respectively. CONCLUSIONS AND IMPLICATIONS: We developed robust ML models to predict death within 30 or 180 d after discharge to an SNF for post-acute care. Although further calibration and external validation are necessary, availability of such tools may facilitate risk-stratified, patient-centered care and prepare clinicians for discussions of care goals.