Sara Seri - Capstone #3 Research Package

Title:
Predictive Analytics for Healthcare Cost Reduction: Using Artificial Intelligence to Identify Preventable High-Cost Patients Before Crisis Care

Study design:
- Source: AHRQ MEPS HC-252 Panel 27 longitudinal file (2022-2023)
- Predictors: 2022 information only
- Outcome: weighted top-decile total healthcare expenditure in 2023
- Models: Logistic Regression, Decision Tree, Random Forest, XGBoost
- Development splits: 70/30, 80/20, and 95/5
- Sealed final validation: 10%

Principal result:
Random Forest was selected. On sealed validation, ROC-AUC = 0.8171, PR-AUC = 0.4429, recall = 0.6042, precision = 0.4715, and top-decile lift = 4.3130.

Key files:
- Sara_Seri_Capstone3_Healthcare_Cost_Reduction.docx: complete thesis draft
- Sara_Seri_Capstone3_Healthcare_Cost_Reduction.pdf: rendered thesis
- Capstone3_Healthcare_Cost_Results.xlsx: executive dashboard and results tables
- analytic_dataset.csv: cleaned research dataset
- model_performance_all_splits.csv: model comparison
- sealed_validation_metrics.csv: final validation results
- feature_importance.csv: final predictor importance
- stress_test_scenarios.csv: scenario-based stress test
- figure_*.png: publication-quality figures

Important limitation:
The model predicts high-cost status but does not prove that a particular intervention will prevent costs. Local validation, clinical review, fairness analysis, governance, and prospective evaluation are required before operational use.
