MERGE: A model for multi-input biomedical federated learning.

Area of Science:

• Biomedical informatics
• Artificial intelligence in healthcare
• Machine learning for medical imaging

Background:

• Deep learning (DL) is crucial for analyzing medical images but often neglects valuable tabular patient data.
• Large datasets for DL require data pooling, posing significant privacy risks.
• Federated learning (FL) offers a privacy-preserving solution by training models locally.

Purpose of the Study:

• To develop and evaluate a federated multi-input architecture integrating medical images and tabular data.
• To enhance AI model performance and generalizability while ensuring data privacy.
• To demonstrate the methodology's effectiveness in real-world biomedical applications.

Main Methods:

• Implemented a federated multi-input neural network architecture.
• Integrated both imaging and tabular patient data within the federated learning framework.
• Validated the approach on COVID-19 prognosis and Alzheimer's disease patient stratification tasks.

Main Results:

• The federated multi-input model achieved higher accuracy and F1 scores compared to single-input models.
• The proposed method demonstrated improved generalizability over non-federated approaches.
• Privacy was maintained by training models locally across institutions.

Conclusions:

• Federated learning combined with multi-modal data (images and tabular) significantly enhances AI performance in biomedical tasks.
• This approach effectively addresses privacy concerns associated with large-scale medical data analysis.
• The methodology shows promise for advancing AI applications in diagnostics and patient stratification.