Healthcare data is highly siloed due to strict privacy regulations. This limits our ability to train large, robust machine learning models because we cannot centralize patient records.
To address this challenge, I designed a **Federated Learning Framework** enabling collaborative training across 5 distributed client nodes without transmitting any raw medical insurance records.
The Federated Average (FedAvg) Protocol
The architecture follows a star topology:
1. **Clients (Hospital Nodes)**: Train a local **SGDRegressor** model on local data.
2. **Aggregation Server**: Coordinates training rounds, aggregates client weights using the **FedAvg** algorithm, and broadcasts the updated model back to clients.
Handling Non-IID Data
Data distributions across hospital sites are naturally heterogeneous (non-IID). We resolved this by weighting each client's update based on its local sample size, ensuring larger datasets had proportional influence on global model convergence while preserving patient privacy.