07 · Personal & Academic Project
Hospital Readmission Risk
Turning fragmented EHR history into an actionable early-warning signal
Business value
The model reached 0.82 ROC-AUC and produced a ranked high-risk list for care-coordination prioritization.
01 · Business problem
Patient data was spread across 16 relational tables with duplicate IDs, overlapping dates, and inconsistent care episodes.
02 · My approach
I resolved identity and time alignment first, engineered clinically meaningful features, and used leakage-safe evaluation.
03 · What I built
PostgreSQL integration, keyed joins, date alignment, clinically informed imputation, 22 features, stratified validation, and Streamlit.
What changed my thinking
The corrections, tradeoffsand decisions behind it.
- 01
Integration was harder than modeling
The same patient appeared across multiple tables and care episodes with conflicting timing.
- 02
Leakage changed the evaluation
Future information had to be removed before performance could be trusted.
- 03
A score was not enough
The final output emphasized ranked patients and understandable risk factors.