Predicting Early Unplanned Readmission to Critical Care Units Using Machine Learning and Metaheuristic Hyperparameter Optimization: Evidence from a Chilean Public Hospital

Background: Early unplanned readmission to critical care units is associated with adverse outcomes and consumes scarce capacity twice, yet the decision to transfer a patient out of the unit is still made without objective criteria. Machine learning may support this decision, but metaheuristic hyperparameter search remains largely unexplored here. Methods: We analyzed 371 patients discharged from the critical care unit of a Chilean public hospital, of whom 41 were readmitted within seven days of transfer to a general ward. Five models (logistic regression, support vector classification, decision tree, random forest, and XGBoost) were trained on clinical and nursing variables available at transfer, using nested stratified five-fold cross-validation with preprocessing fitted within the training folds. Bayesian optimization and a genetic algorithm were compared under identical folds. Results: Under nested cross-validation, discrimination was moderate (AUC 53.3–75.2%), and no single model performed best across all metrics. Logistic regression achieved the highest AUC (75.2%) with the lowest variability across folds, followed by XGBoost (75.0%), whose results varied markedly across folds. The genetic algorithm yielded higher AUC than Bayesian optimization for most tree-based models. Days in the critical care unit, age, and unit of origin ranked highest in impurity-based importance. Conclusions: This proof-of-concept study suggests that routinely collected clinical and nursing information available at transfer may inform early readmission prediction. As the sample was case-enriched and derived from a single center, recalibration to the natural readmission prevalence and prospective external validation are required before clinical implementation.