Predicting Adverse Drug Reactions on Distributed Health Data using Federated Learning

Area of Science:

• Health Informatics
• Machine Learning
• Pharmacovigilance

Background:

• Predicting adverse drug reactions (ADRs) using electronic health data faces challenges like data scarcity at individual sites, integration costs, and privacy concerns.
• Centralized databases for ADR prediction are often impractical due to the sensitive nature of patient data.

Purpose of the Study:

• To develop a federated learning framework for predicting ADRs from distributed electronic health data.
• To propose novel local model aggregation methods to enhance global ADR prediction model performance.

Main Methods:

• Implemented a federated learning framework to train a global ADR prediction model on decentralized health data.
• Developed and evaluated two new methods for aggregating local models to improve predictive accuracy.
• Conducted experiments on real-world health data from 1 million patients.

Main Results:

• The federated learning approach achieved performance comparable to centralized learning methods.
• The proposed framework outperformed localized learning models for predicting two types of ADRs.
• Novel aggregation methods demonstrated superior precision, recall, and accuracy compared to state-of-the-art techniques across varying data distributions.

Conclusions:

• Federated learning offers a viable solution for ADR prediction using distributed electronic health data, overcoming limitations of centralized approaches.
• The proposed aggregation methods significantly enhance the predictive power of federated models for ADR detection.
• This framework provides a privacy-preserving and efficient method for large-scale pharmacovigilance.