The Federated Sample Weight Model: A New Strategy for Precision Medicine

Abstract

As artificial intelligence advances in the medical field, machine learning has shown significant promise in precision medicine. However, the development of accurate models is often constrained by the need for centralized training on large datasets, which is limited by privacy concerns and data-sharing restrictions between medical institutions. Additionally, uneven data distribution and skewed datasets can lead to biased training, impacting model performance. To address these challenges, we introduce FedSam, a novel federated sample weight approach designed to optimize federated learning models and enhance cardiovascular disease prediction. FedSam, integrated with federated learning models, enables the training of samples from multiple medical institutions while preserving privacy. By applying dynamic weights to local samples within the federated learning framework, FedSam improves the identification of biased samples and facilitates the effective sharing of medical resources. Experimental results using a cross-sectional study from a tertiary hospital in China demonstrate that FedSam achieves superior performance in patient condition identification compared to four conventional federated learning methods. The findings highlight that FedSam enhances prediction capabilities and increases the robustness of federated learning models, particularly in scenarios with both balanced and skewed label distributions. This work represents a significant step towards leveraging federated learning to advance precision medicine in cardiovascular disease management.

Ähnliche Arbeiten