Healthcare Analytics (ASA DataFest)

Built over a DataFest weekend on about 700K healthcare events. It won Best Visualization.

The data goes into a DuckDB warehouse, and a Streamlit dashboard sits on top. High-utilizer patients are segmented with k-means on PCA-reduced features, and care journeys are shown as networks so you can see where patients actually move through the system.

The predictive half asks whether a patient’s follow-up will be delayed, at four time horizons. XGBoost and CatBoost classifiers on leakage-aware cohort features, validated with patient-grouped splits so no patient appears on both sides, reach an AUCPR of about 0.85.