Among 3,933 heart failure patients discharged from New South Wales hospitals, 45.2% experienced emergency readmission and 12.4% died within 180 days. Machine learning models trained on electronic health records — incorporating clinical, laboratory, medication, and text-derived variables — achieved moderate readmission prediction (AUC 0.70) and strong mortality prediction (AUC 0.84). Key readmission drivers included prior healthcare utilization, fall risk, polyproblem burden, age, and renal impairment. Mortality was most powerfully predicted by red blood cell distribution width (RDW), blood urea nitrogen, age, and low systolic blood pressure — a mechanistically coherent signature of cardiorenal-anemic syndrome.
The near-50% readmission rate within six months reinforces what clinicians already know: hospital discharge marks the beginning of a critical vulnerability window, not a resolution. What distinguishes this work is the Shapley-value explainability layer, which moves beyond black-box risk scores toward individualized risk narratives — a meaningful step for clinical adoption. The paradoxical finding that fewer discharge medications correlated with higher readmission but lower mortality likely reflects end-of-life deprescribing patterns, a nuance worth scrutinizing carefully. Limitations include retrospective design, single-state data, and moderate readmission discrimination (AUC 0.70) that may limit real-world utility for triage decisions. The calibration slope of 1.30 for readmission also signals some overconfidence in predicted probabilities. As a preprint not yet peer-reviewed, these findings require independent validation before clinical deployment. Incremental rather than paradigm-shifting, but the explainability framework offers a practical template for responsible AI integration in post-discharge care.
