ReadmitIQ
Can a hospital see a 30-day readmission coming from the data it already collects? A two-dataset study where most of the work was turning raw clinical records into an answerable question.
scikit-learnXGBoostMIMIC-IIIFeature engineering
101,766
Encounters
6M+
Lab events merged
3
Models compared
What I built
Framed 30-day readmission as a prediction problem on two public datasets: a 10,000-patient MIMIC-III subset (seven tables merged, 6M+ lab events) and 101,766 encounters from the UCI Diabetes dataset.
Spent most of the effort making the data answerable — deriving the readmission label from raw admission and discharge timestamps, then engineering length of stay, admission frequency, a composite risk score and 23 per-drug medication-change encodings.
Compared Logistic Regression, Random Forest and XGBoost on each dataset, taking the 5.5–11.2% class imbalance seriously with class weighting and stratified splits, and judging models on ROC AUC and F1 rather than raw accuracy.