Federated Learning for Healthcare Insurance

Privacy-Preserving Premium Modeling and Fraud Detection

Overview

Digital healthcare insurance systems generate vast amounts of demographic, behavioral, biometric, and claims-related data across geographically distributed institutions. Centralizing this data for machine learning introduces significant privacy, regulatory (HIPAA), and data-governance risks.

This project implements a Federated Learning (FL) framework using the Flower library that trains:

• Premium prediction models — personalized actuarial modeling from distributed data
• Fraud detection models — proactive identification of anomalous claims

Models are trained across 5 simulated distributed nodes without any raw data ever leaving a local node.

---

Key Results

Metric	Centralized	Federated (5 nodes)	Privacy Cost
Premium RMSE	~$4,800	~$4,950	+2–3%
Fraud AUC-ROC	~0.921	~0.897	−2–3%
Raw data shared	100%	0%	Gradient updates only

Finding: Federated models achieve within 3–4% of centralized performance while sharing zero raw patient records — making this approach directly applicable to HIPAA-regulated workers' compensation and healthcare insurance environments.

---

Architecture

Global Model (Server — FedAvg)
      │
      ├── Distribute weights ──► Node 1 (Region: North)
      ├── Distribute weights ──► Node 2 (Region: South)
      ├── Distribute weights ──► Node 3 (Region: East)
      ├── Distribute weights ──► Node 4 (Region: West)
      └── Distribute weights ──► Node 5 (Region: Central)
                                         │
                      ◄──────────────────┘
              Aggregate gradient updates (NO raw data transmitted)

Each node receives a non-IID partition — different demographic distributions per region — which is the realistic challenge in federated healthcare settings.

Privacy Mechanism

We implement Differential Privacy (DP-FedAvg) with the Gaussian mechanism:

$$\tilde{\Delta w} = \Delta w + \mathcal{N}(0, \sigma^2 C^2 \mathbf{I})$$

where C is the L2 sensitivity (gradient clipping norm) and σ controls the privacy-accuracy tradeoff. The notebook evaluates σ ∈ {0.1, 0.3, 0.5, 1.0} and plots the resulting privacy budget vs. accuracy curve.

---

Project Structure

Federated/
├── FEDERATED.ipynb              # Main research notebook (run this)
├── requirements.txt             # Pinned dependencies
├── README.md                    # This file
├── CHANGELOG.md                 # Version history
├── CONTRIBUTING.md              # Contribution guide
├── LICENSE                      # MIT License
├── DATA/
│   ├── Medical_insurance.csv    # Original dataset
│   └── insurance_federated.csv  # Generated synthetic dataset (by notebook)
└── outputs/                     # Generated plots and results
    ├── 01_eda.png
    ├── 02_comparison.png
    ├── 03_dp_tradeoff.png
    └── 04_final_summary.png

---

Notebook Structure

Section	Description
1. Environment Setup	Dependencies, imports, version checks
2. Dataset Generation	Synthetic 5,000-record insurance dataset with realistic distributions
3. EDA	Distribution plots, correlation heatmap, fraud rate by region
4. Centralized Baseline	GBM models trained on pooled data (performance ceiling)
5. FL Architecture	Non-IID partitioning, FedAvg implementation
6. Federated Premium	20-round federated training for premium regression
7. Federated Fraud	20-round federated training for fraud classification
8. Comparison	Convergence curves, per-node evaluation, summary charts
9. Differential Privacy	DP-FedAvg, privacy-accuracy tradeoff analysis
10. Results	Final metrics summary, key findings
11. References	Academic citations

---

Installation

git clone https://github.com/ramkalapala/Federated.git
cd Federated
python -m venv venv
source venv/bin/activate      # Windows: venv\Scripts\activate
pip install -r requirements.txt
jupyter notebook FEDERATED.ipynb

---

Requirements

See requirements.txt. Core dependencies:

flwr>=1.5.0
scikit-learn>=1.3.0
numpy>=1.24.0
pandas>=2.0.0
matplotlib>=3.7.0
seaborn>=0.12.0
jupyter>=1.0.0
nbformat>=5.9.0

---

Motivation & Real-World Applicability

This research was motivated by the operational challenges in workers' compensation claims processing — specifically the tension between:

1. The need for accurate predictive models across distributed provider/insurer networks
2. Strict data privacy requirements (HIPAA, California CMIA, workers' comp regulations)
3. The impracticality of centralizing PHI across organizational boundaries

The federated approach demonstrated here directly addresses this gap. The architecture — distributed training with FedAvg aggregation, DP noise injection at the gradient level, and human-in-the-loop routing at confidence boundaries — maps directly onto the multi-node structure of real insurance data ecosystems.

---

Citation

If you use this work, please cite:

@software{kalapala2024federated,
  author    = {Kalapala, Ramprakash},
  title     = {Federated Learning for Healthcare Insurance: Privacy-Preserving Premium Modeling and Fraud Detection},
  year      = {2024},
  publisher = {GitHub},
  url       = {https://github.com/ramkalapala/Federated}
}

---

References

1. McMahan et al. (2017). Communication-Efficient Learning of Deep Networks from Decentralized Data. AISTATS 2017. arXiv:1602.05629
2. Dwork & Roth (2014). The Algorithmic Foundations of Differential Privacy.
3. Beutel et al. (2020). Flower: A Friendly Federated Learning Research Framework. arXiv:2007.14390
4. Rieke et al. (2020). The Future of Digital Health with Federated Learning. npj Digital Medicine.
5. Kairouz et al. (2021). Advances and Open Problems in Federated Learning. Foundations and Trends in ML.

---

License

MIT License — see LICENSE

Author

Ramprakash Kalapala — Senior Cloud Solution Architect
Focus: AI/ML systems for regulated healthcare and insurance environments