Health Insurance Fraud Detection

A comprehensive machine learning project for detecting and preventing health insurance fraud using advanced classification models and domain-specific heuristic triggers.

📋 Project Overview

This project develops a fraud detection system that uses machine learning models to identify potentially fraudulent health insurance claims. The system combines traditional ML algorithms with domain-specific fraud detection triggers to provide a robust, comprehensive fraud detection solution.

Key Objectives

• Red flag claims for possible fraud and further investigation
• Reduce losses and increase profitability for insurers/reinsurers
• Reduce premiums for customers' insurance coverage
• Provide competitive advantage for insurers/reinsurers

Key Features

• Multiple ML models: Decision Tree, Random Forest, XGBoost, GLM, Naive Bayes, GBM
• Imbalanced dataset handling: ADASYN, SMOTE, MWMOTE, ROSE
• Domain-specific fraud triggers (8 types)
• Comprehensive performance evaluation with ROC-AUC, PR-AUC, and classification metrics

📁 Project Structure

CAS-Project-Health-Insurance-Fraud-Detection/
├── data/
│   ├── raw/              # Original, unprocessed data
│   ├── interim/          # Intermediate transformed data
│   └── external/         # External data sources (triggers, hospital lists)
├── src/
│   ├── features/         # Feature engineering & trigger functions
├── results/              # Model outputs and results
├── docs/                 # Additional documentation
├── main.Rmd              # Main markdown file which contains the code
└── README.md

🚀 Getting Started

Prerequisites

R Version: 4.0 or higher

Required R Packages:

# Data manipulation
library(readxl)
library(janitor)
library(dplyr)
library(tidyverse)
library(lubridate)

# Machine Learning
library(caret)
library(rpart)
library(rpart.plot)
library(randomForest)
library(xgboost)
library(gbm)
library(e1071)
library(naivebayes)
library(glmnet)

# Imbalanced data handling
library(smotefamily)
library(imbalance)
library(ROSE)

# Evaluation & Visualization
library(pROC)
library(PRROC)
library(ROCR)
library(ggplot2)
library(kableExtra)

# Utilities
library(psych)
library(bruceR)
library(caTools)
library(mlbench)

Installation

1. Clone the repository:

git clone https://github.com/RohanYashraj/CAS-Project-Health-Insurance-Fraud-Detection.git
cd CAS-Project-Health-Insurance-Fraud-Detection

2. Install R dependencies:

# Run this in R console
install.packages(c("readxl", "janitor", "dplyr", "tidyverse", "lubridate",
                   "caret", "rpart", "rpart.plot", "randomForest", "xgboost",
                   "gbm", "e1071", "naivebayes", "glmnet", "smotefamily",
                   "imbalance", "ROSE", "pROC", "PRROC", "ROCR", "ggplot2",
                   "kableExtra", "psych", "bruceR", "caTools", "mlbench"))

Usage

1. Prepare Data: Place your cleaned input data in data/raw/cleaned_input_data.rds
2. Run Main Analysis: Open and knit notebooks/02-modeling/final_code.Rmd
3. Review Results: Check results/ folder for outputs

Main analysis workflow:

• Data preprocessing and feature engineering
• Application of fraud detection triggers
• Dataset balancing using multiple techniques
• Model training across 6 algorithms
• Performance evaluation and comparison

🔍 Fraud Detection Triggers

The project implements 8 domain-specific fraud triggers:

1. Claim Amount Trigger: Flags claims exceeding expected package amounts
2. Hospital Days Trigger: Identifies excessive length of stay
3. Age Trigger: Flags procedures inappropriate for patient's age group
4. Gender Trigger: Detects gender-specific procedure mismatches
5. Claim Count Trigger: Identifies excessive claim frequency
6. Close Proximity Trigger: Flags claims filed too soon after policy commencement
7. Treatment Date Validity: Ensures treatments within policy coverage periods
8. Claim Reported Delay: Identifies unusual reporting delays

🤖 Machine Learning Models

Algorithms Evaluated

• Decision Tree: Interpretable rule-based classifier
• Random Forest: Ensemble of decision trees
• XGBoost: Gradient boosting framework
• GLM: Generalized Linear Model (Logistic Regression)
• Naive Bayes: Probabilistic classifier
• GBM: Generalized Boosted Regression Model

Imbalanced Data Handling

Each model is evaluated with and without four different balancing techniques:

• ADASYN: Adaptive Synthetic Sampling
• SMOTE: Synthetic Minority Oversampling Technique
• MWMOTE: Majority Weighted Minority Oversampling Technique
• ROSE: Random Over-Sampling Examples

📊 Performance Metrics

Models are evaluated using:

• Accuracy: Overall classification accuracy
• Sensitivity (Recall): True positive rate
• Specificity: True negative rate
• Precision: Positive predictive value
• F1-Score: Harmonic mean of precision and recall
• F2-Score: Weighted F-score prioritizing recall
• ROC-AUC: Area under ROC curve
• PR-AUC: Area under Precision-Recall curve

📈 Results

Comprehensive results comparing all model configurations are available in:

• results/final_output.xlsx: Detailed performance metrics
• results/results.rds: R data format

👥 Authors

SSSIHL-CAS Project Team

📝 License

This project is proprietary. To access, please send request to Authors of the article.

🙏 Acknowledgments

• SSSIHL for institutional support
• Insurance industry partners for data and domain expertise

📧 Contact

For questions or access requests, please contact the project authors.

---

Note: Due to the sensitive nature of fraud detection data, the data has been masked.