Prediction of unplanned readmission in older adults during hospital stay: a data-driven approach to support care transition

PURPOSE/AIMS: Accurately predicting 30-day unplanned readmission in older adults is critical for improving care transitions and reducing preventable hospitalizations. Most existing models rely only on data available at discharge, limiting early intervention. This study aimed to develop and evaluate machine learning models to predict 30-day unplanned readmissions using information from the first 48 h of hospitalization, and to compare their performance with discharge-based models. METHODS: We analyzed data from 14,140 older adult admissions to a quaternary hospital. We built ML algorithms to predict 30-day unplanned readmission at 48 h after admission and at discharge [Logistic Regression, Random Forest, Light Gradient Boosting Machine (LGBM), and others] using administrative data and clinical information extracted from the electronic health record. Model's performance was assessed using confusion matrix and discrimination metrics. RESULTS: The Light Gradient Boosting Machine demonstrated the best performance at both 48 h (Sensitivity = 79%, Specificity = 59%, AUC-ROC = 0.75) and discharge (Sensitivity = 74%, Specificity = 65%, AUC-ROC = 0.76). Discriminative ability at 48 h was comparable to discharge (DeLong test p-value = 0.17). The application of both, the 48-hour and discharge models, could potentially predict 82% of readmissions. CONCLUSION: ML algorithms using data commonly obtained in electronic health records showed fair performance to predict 30-unplanned readmission in older adults using data from the first 48 h after admission, which was comparable to performance at discharge. If corroborated by external validation studies, these findings create opportunities to design and implement risk mitigation interventions during hospital stay.