HealthFraudMLChain

Enhancing Healthcare Insurance Fraud Detection and Prevention with a Machine Learning and Blockchain-Based Approach

M.Sc. Dissertation, Department of Mathematics, National Institute of Technology Patna, 2024

Live Demo: healthcare-fraud-detection.pages.dev

Overview

This repository contains the complete research pipeline for healthcare insurance fraud detection using an ensemble of Optuna-tuned machine learning classifiers with blockchain-secured audit trails and ECIES encryption. The system aggregates 558,211 Medicare claims into 5,410 provider-level records with 190 engineered features, achieving F1 = 0.7345 with zero data leakage under 10-fold stratified cross-validation.

Key Results

Metric	Value	95% Bootstrap CI
F1 Score	0.7345	[0.7118, 0.7715]
Precision	73.7%	[69.8%, 77.5%]
Recall	74.7%	[70.9%, 78.4%]
MCC	0.7151	[0.6818, 0.7474]
ROC-AUC	0.9587	[0.9477, 0.9686]
PR-AUC	0.8114	[0.7801, 0.8406]

Weighted ensemble: LightGBM (39.6%), CatBoost (32.0%), XGBoost (28.2%) at fixed threshold t = 0.444.

Statistical significance: Friedman test p = 0.00089; all pairwise Wilcoxon comparisons significant after Holm correction.

Features

• ML Pipeline: Five gradient-boosted classifiers (XGBoost, LightGBM, CatBoost, GradientBoosting, RandomForest) with Optuna TPE hyperparameter optimization (60 trials per model)
• Leakage-Free Evaluation: All feature engineering (z-scores, percentiles, IsolationForest, physician statistics) computed strictly within each CV fold
• Explainability: Dual SHAP (global) + LIME (local) interpretation identifying deductible patterns, reimbursement anomalies, and claim duration signals as top fraud indicators
• Blockchain Audit: Custom SHA-256 blockchain with Merkle tree integrity verification, PoW (difficulty 2), and SQLite persistence (110 blocks, 5,411 records)
• ECIES Encryption: secp256k1 + HKDF-SHA256 + AES-256-GCM field-level encryption for provider PII protection
• Statistical Rigor: Friedman + Wilcoxon (Holm-corrected) non-parametric tests, Cohen's d effect sizes, 2000-iteration bootstrap confidence intervals
• Web Prototype: Interactive Next.js static site deployed on Cloudflare Pages with browser-side ECIES demo

Repository Structure

corrected_pipeline/
  01_preprocess.py              # Data preprocessing and feature engineering
  02_train_evaluate.py          # Initial model training
  04_optuna_tuning.py           # Optuna hyperparameter optimization
  15_definitive_final.py        # Definitive leakage-free evaluation (main script)
  16_regenerate_all_figures.py  # Figure generation
  17_multi_node_demo.py         # Multi-node blockchain consensus demo
  export_static_data.py         # Export data for web frontend
  blockchain/
    chain.py                    # SHA-256 blockchain with Merkle trees
    ecies_cipher.py             # ECIES encryption (secp256k1 + AES-256-GCM)
  results/
    definitive_final_results.json  # All model metrics and statistical tests
    figures/                       # 15 research figures (PNG + PDF)
  thesis_latex/
    main.tex                    # Complete thesis LaTeX source
    chapters/                   # Ch1-Ch5 chapter files
Dataset/                        # Kaggle rohitrox Medicare claims data
thesis.pdf                      # Compiled thesis PDF

Quick Start

# Clone
git clone https://github.com/ascender1729/HealthFraudMLChain.git
cd HealthFraudMLChain/corrected_pipeline

# Install dependencies
pip install -r requirements.txt

# Run the definitive evaluation (requires Dataset/ with Medicare CSVs)
python 15_definitive_final.py

# Run blockchain demo
python 06_blockchain_demo.py

# Multi-node consensus validation
python 17_multi_node_demo.py

Web Prototype

The interactive demo is deployed at healthcare-fraud-detection.pages.dev with:

• Provider Analysis - Look up any provider's fraud risk with SHAP/LIME explanations
• Research Results - All 15 figures, statistical tests, bootstrap CIs
• Blockchain Explorer - Inspect the 110-block audit chain
• ECIES Demo - Live browser-side encrypt/decrypt using Web Crypto API

Tools and Technologies

Category	Technologies
ML Pipeline	Python, scikit-learn, XGBoost, LightGBM, CatBoost, Optuna, SHAP, LIME
Blockchain	SHA-256, Merkle Trees, Proof-of-Work, SQLite
Encryption	ECIES (secp256k1 + HKDF-SHA256 + AES-256-GCM)
Statistics	Friedman test, Wilcoxon signed-rank, Holm correction, Bootstrap CI
Web	Next.js 14, TypeScript, Tailwind CSS, Magic UI, Cloudflare Pages
Data	Kaggle rohitrox Medicare claims (558K claims, 5,410 providers)

License

This work is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License.

Acknowledgments

• Dr. Rajesh Kumar Sinha for guidance and supervision
• National Institute of Technology Patna for research facilities

---

Citation

If you use this project in your research, please cite:

@thesis{dubasi2024_healthfraudmlchain,
  author  = {Dubasi, Pavan Kumar},
  title   = {Enhancing Healthcare Insurance Fraud Detection and Prevention
             with a Machine Learning and Blockchain-Based Approach},
  school  = {National Institute of Technology Patna},
  year    = {2024},
  type    = {Integrated M.Sc. Mathematics Dissertation},
  advisor = {Dr. Rajesh Kumar Sinha},
  url     = {https://github.com/ascender1729/HealthFraudMLChain}
}

Author

Dubasi Pavan Kumar

• Website: dubasipavankumar.com
• LinkedIn: in/im-pavankumar
• ORCID: 0009-0006-1060-4598