Performance Analysis of Machine Learning Algorithms in Identifying the Risk of Readmission for Inpatients at the Langsa Regional General Hospital

Abstract

This paper presents a preliminary descriptive analysis and an evaluation design for Random Forest and Extreme Gradient Boosting (XGBoost) to identify unplanned inpatient readmission risk at RSUD Langsa. Aggregate counts were extracted from the research proposal for index discharges between July 16, 2024 and June 16, 2026. The screened cohort comprised 34,122 episodes from 25,821 patients, including 2,870 unplanned readmissions occurring more than 24 hours and up to 30 days after discharge. The readmission proportion was 8.41%, compared with 13.63% among chronic-disease episodes and 6.94% among nonchronic episodes. Patients aged 60–74 years had a proportion of 13.78%. An analytical all-negative classifier would achieve 91.59% accuracy while detecting no readmissions. This supports evaluating precision–recall performance, recall, F2-score, and calibration rather than accuracy alone. The planned experiment uses patient-level splitting, nested cross-validation, and comparisons of unadjusted training, SMOTE-NC, and class weighting. As trained-model outputs and individual predictions are unavailable in the proposal, empirical superiority of either algorithm cannot yet be established. This stage contributes a descriptive outcome profile and an evaluation framework that accounts for class imbalance and predictor availability at discharge.