Machine Learning Prediction of 90-Day Readmission and Mortality in General Internal Medicine: A Retrospective Cohort Study

Background/Objectives: Predicting post-discharge outcomes in general internal medicine (GIM) is essential for targeted intervention. This study developed machine learning (ML) models to predict 90-day hospital readmission and 90-day all-cause post-discharge mortality using routine clinical data. Methods: This retrospective cohort study included 935 adult admissions to a tertiary GIM department (January 2022–December 2023), of which 881 were retained after data quality verification. Six ML algorithms—Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine (SVM), Neural Network, and Voting Ensemble—were evaluated across admission-only (Tier 1) and discharge-augmented (Tier 2) feature sets using Synthetic Minority Over-sampling Technique (SMOTE), SelectKBest feature selection, and five-fold cross-validated hyperparameter tuning. Results: The 90-day readmission and mortality rates were 33.4% and 4.8%, respectively. Mortality models substantially outperformed readmission models (best admission-model AUC 0.821 for the Voting Ensemble vs. 0.645 for Logistic Regression in readmission; best discharge-model mortality AUC 0.828 for Logistic Regression, 95% CI: 0.694–0.930). Discharge models did not improve over admission models for either outcome. Key mortality predictors included code status, albumin level, functional dependence, and age. Conclusions: Admission-available clinical variables enabled good discrimination of 90-day mortality but only modest readmission prediction. Given the low events-per-variable ratio and single-centre design, these findings should be considered hypothesis-generating and require external validation before clinical implementation.