PRISM-FL–A Privacy-Aware Decision Support System for Enhanced Medical Diagnostics Based on Federated Learning
Barbara Pȩkala, Przemysław Grzegorzewski, Jarosław Szkoła, Dawid Kosior · IEEE Access · 2025
This study tackles the challenge of building effective diagnostic models while preserving patient data privacy, a critical issue in medical applications like breast cancer diagnosis. Individual healthcare institutions often lack sufficient or high-quality data to train reliable machine learning models, and privacy regulations prevent data sharing. To address this, we use federated learning, a technique that allows multiple organizations to collaboratively train a model without sharing local data. Our focus is on horizontal federated learning, where each participant trains a local model on its own data. These models are updated locally, then aggregated periodically to improve the global model’s performance. Aggregation typically relies on weighted averaging, with different strategies for assigning weights. We enhance this process by incorporating uncertainty-aware aggregation using the Choquet integral, a method for information fusion. And also by integrating uncertainty measures such as entropy, we aim to improve both model performance and robustness. We also investigate how local model parameters are affected by their performance and by uncertainties present in the data.