Readmission prediction .
A model result needs its tradeoffs beside it.
Concept illustration / A view of the idea, not a product screenshot.
01 / The context
The problem
worth working on.
Predicting 30-day readmission is an imbalanced classification problem. A single accuracy score can hide missed cases, excessive false positives, and uneven performance across groups.
02 / The approach
How the pieces
come together.
- 01
Completed the ML lifecycle from preprocessing and feature selection through model comparison, threshold tuning, and reproducible artifact export.
- 02
Compared baseline, tuned, and PCA-based candidates. The repository reports a Random Forest with PCA as its selected model.
- 03
Paired the exported pipeline with a FastAPI interface and Streamlit application, while documenting explainability and fairness analysis.
03 / A key decision
The detail
that shapes the system.
Report precision alongside recall. At the selected threshold, the repository reports 71.64% recall and 15.00% precision; higher recall comes with substantial false positives.
04 / The result
What came out of it.
A completed capstone with a reproducible inference pipeline. Reported test AUC-ROC is 0.6446, below the original 0.75 target. The reported racial recall gap also exceeded its target.
Scope & limitations
Academic research on historical data, not a clinically validated decision tool. Metrics are repository-reported results, not an independent evaluation; external validation would be necessary before practical use.