Explainable Random Forest Prediction of 30-Day Hospital Readmission

Early return to hospital after discharge is a widely used signal of care quality and a major driver of avoidable cost. This paper reports an explainable machine-learning study that estimates whether a diabetic patient will be readmitted within 30 days, using the public Diabetes 130-US Hospitals data. Keeping only the first encounter of every patient left 69,987 records, of which 5,143 (7.35%) were early readmissions. A single Random Forest classifier with 150 trees and balanced class weights was fitted on 55,989 records and checked on 13,998 unseen ones. It reached 85.1% accuracy, yet precision was 0.174, recall 0.276, F1-score 0.214 and ROC-AUC 0.665, which shows how misleading accuracy becomes under heavy imbalance. Permutation importance ranked the count of earlier inpatient stays first, followed by discharge to skilled nursing and rehabilitation facilities. A 20-feature subset kept almost all of the performance, and a threshold study exposed the precision–recall trade-off that a hospital must manage. The model is best seen as a decision-support aid that complements, and never replaces, clinical judgement.