Trelegy Ellipta — Copay Fraud Detection with Isolation Forest

Product: Trelegy Ellipta (GSK) — Fluticasone furoate / Umeclidinium / Vilanterol
Method: Hybrid Rules + Isolation Forest + SHAP Explainability
Use Case: Detecting copay card fraud in pharmaceutical patient assistance programs

🏗️ Project Overview

Why Trelegy?

• High-value drug (~$600-700/month retail) → high copay card value per claim
• Chronic daily use → monthly refills create recurring fraud opportunities
• Fixed 30-day supply (exactly 30 blisters per inhaler) → easy to detect refill anomalies
• Multiple assistance programs → card stacking opportunities
• Even 2% fraud leakage on Trelegy = ~$160M/year in losses

📊 Fraud Types Detected

🏛️ Architecture

Raw Copay Claims
      │
      ▼
┌─────────────────┐
│ Phase 1: Rules  │ ← 10 Trelegy-specific hard rules
│ (Hard Flags)    │   (early refill, impossible qty, govt insurance, etc.)
└────────┬────────┘
         ▼
┌─────────────────┐
│ Phase 2: IF     │ ← Isolation Forest trained on rule-clean data
│ (Anomaly Detect)│   (catches unknown/novel fraud patterns)
└────────┬────────┘
         ▼
┌─────────────────┐
│ Phase 3: Score  │ ← Combined priority = 50% IF + 30% rule severity + 20% rule flag
│ (Priority Rank) │   Risk tiers: Low / Medium / High / Critical
└────────┬────────┘
         ▼
┌─────────────────┐
│ Phase 4: SHAP   │ ← TreeExplainer → "Why was this claim flagged?"
│ (Explainability)│   Top features driving each anomaly score
└────────┬────────┘
         ▼
┌─────────────────┐
│ Phase 5: Eval   │ ← AUPRC, AUROC, F1, Precision@K, detection by fraud type
│ (Reporting)     │   + 5 publication-quality visualizations
└─────────────────┘

📁 Project Structure

trelegy_copay_fraud/
├── README.md                          # This file
├── run_all.py                         # 🚀 Master script — run this
├── generate_synthetic_data.py         # Synthetic Trelegy copay claims generator
├── feature_engineering.py             # Feature engineering pipeline (48 features)
├── fraud_detection_pipeline.py        # Full detection pipeline (rules + IF + SHAP + eval)
├── requirements.txt                   # Python dependencies
├── data/                              # Generated synthetic data
│   ├── trelegy_copay_claims.csv       # Main claims dataset
│   ├── patient_master.csv             # Patient demographics
│   ├── pharmacy_master.csv            # Pharmacy reference
│   └── prescriber_master.csv          # Prescriber reference
└── results/                           # Pipeline outputs
    ├── investigation_queue_top500.csv # Top 500 flagged claims for review
    ├── scored_claims_full.csv         # All claims with scores & risk tiers
    ├── feature_importance_shap.csv    # SHAP feature importance ranking
    ├── metrics.json                   # Evaluation metrics
    ├── 01_evaluation_metrics.png      # ROC, PR curve, confusion matrix
    ├── 02_shap_summary.png            # SHAP beeswarm plot
    ├── 03_shap_bar_importance.png     # SHAP bar chart
    ├── 04_detection_by_fraud_type.png # Detection rates by fraud type
    ├── 05_risk_tier_distribution.png  # Risk tier analysis
    └── model/
        ├── isolation_forest_model.pkl # Trained IF model
        ├── scaler.pkl                 # StandardScaler
        ├── encoder.pkl                # OrdinalEncoder
        └── feature_names.pkl          # Feature name list

🚀 Quick Start

1. Generate ~50,000 synthetic Trelegy copay claims (with ~3% injected fraud)
2. Engineer 48 features (temporal, behavioral, rolling windows, Trelegy-specific)
3. Apply 10 hard business rules
4. Train Isolation Forest on rule-clean data
5. Compute SHAP explanations for top features
6. Generate evaluation metrics and 5 visualizations
7. Output an investigation queue (top 500 highest-priority claims)

📈 Features (48 Total)

Temporal Features

• days_between_fills — Days since last fill (normal: 28-33 for Trelegy)
• early_refill_flag — Binary: fill before day 23
• days_since_first_fill — Patient tenure in program
• claim_month, claim_dow — Seasonality

Rolling Window Aggregates (7d/30d/90d)

• patient_fill_count_{7,30,90}d — Fill velocity per patient
• patient_copay_spend_{7,30,90}d — Copay card spend per patient
• patient_total_claim_{7,30,90}d — Total claim amount per patient
• pharmacy_claim_count_{30,90}d — Volume per pharmacy
• prescriber_claim_count_{30,90}d — Volume per prescriber

Patient Behavioral

• unique_pharmacies_overall — Number of distinct pharmacies used
• unique_programs_per_patient — Card stacking indicator
• total_fills_per_patient — Lifetime fill count
• avg_days_between_fills — Average refill gap
• std_days_between_fills — Refill pattern consistency
• max_fills_any_30d — Peak fill velocity

Pharmacy & Prescriber

• pharmacy_claims_per_patient_ratio — Ghost fill indicator
• prescriber_specialty_valid — Specialty match flag
• prescriber_total_claims — Volume anomaly

Trelegy-Specific

• ndc_switch_flag — Strength switching (0173-0893 ↔ 0173-0887)
• govt_insurance_flag — Medicare/Medicaid with copay card
• cross_state_fill — Patient state ≠ pharmacy state
• new_patient_burst — >1 fill in first 7 days of enrollment

📐 Evaluation Metrics

• AUPRC — Primary metric (best for imbalanced fraud data)
• Precision@K — Operational metric (investigator queue efficiency)
• F1-Score — Balance of precision and recall
• AUROC — Supplementary (can be misleading at <1% fraud rate)

🔧 Trelegy Product Details

📚 References

• Liu, Ting, Zhou — "Isolation Forest" (IEEE ICDM 2008) — Original algorithm
• Hariri et al. — "Extended Isolation Forest" (2019) — High-dimensional improvement
• Thimonier et al. — arXiv:2312.13896 — IF vs LightGBM fraud benchmark
• Amazon FDB — arXiv:2208.14417 — Feature engineering best practices
• ZS Associates — Pharma Copay Fraud Case Study
• KPMG — Fighting Copay Fraud (PDF)

⚖️ Disclaimer

📄 License