Federated Learning for Healthcare: Class Imbalance Mitigation and Feature Drift Detection.

Area of Science:

• Artificial Intelligence
• Machine Learning
• Healthcare Informatics

Background:

• Federated learning (FL) enables collaborative analysis of decentralized healthcare data, but data quality issues can compromise model reliability and fairness.
• Effective monitoring systems are essential for successful FL implementation in healthcare settings.
• Challenges include label imbalance and feature drift, which can negatively impact model performance.

Purpose of the Study:

• To develop and evaluate a cross-silo FL system for healthcare data using the Flower framework.
• To focus on monitoring metrics and identifying data quality issues like label imbalance and feature drift.
• To assess the impact of data quality monitoring on FL model performance and fairness.

Main Methods:

• Utilized the Flower framework for a cross-silo FL system with a harmonized synthetic dataset from the LETHE project.
• Tested the system on five clients with varying data distributions, including one with significant label imbalance.
• Implemented real-time metric tracking (accuracy, loss, MCC) using Prometheus and Grafana, and integrated feature drift detection (Kolmogorov-Smirnov test).

Main Results:

• The federated model showed competitive performance, though a baseline model initially outperformed it (accuracy: 0.806 vs. 0.754).
• A customized Federated Averaging (FedAvg) algorithm, considering label distribution and dataset size, significantly improved the global model's Matthews Correlation Coefficient (MCC) from 0.111 to 0.349.
• Feature drift detection was successfully integrated, providing visual alerts.

Conclusions:

• Monitoring systems are critical for ensuring the reliability and fairness of FL models in healthcare.
• Customized FL aggregation strategies can mitigate the impact of data heterogeneity and improve model performance.
• Further research is needed to assess customized FedAvg across diverse data distributions and explore advanced privacy techniques.