A privacy-preserving dependable deep federated learning model for identifying new infections from genome sequences

Area of Science:

• Bioinformatics
• Computational Biology
• Machine Learning in Healthcare

Background:

• Traditional molecular identification methods using genome sequences (GSs) face privacy and security challenges due to data transfer.
• Sensitive medical data requires robust privacy-preserving techniques for accurate infection identification.
• Deep federated learning (DFL) offers a promising solution for decentralized, secure analysis of sensitive datasets.

Purpose of the Study:

• To propose a dependable and privacy-preserving DFL-based model for identifying new infections from GSs.
• To address the limitations of traditional methods concerning patient data privacy and security.
• To develop an automated feature selection mechanism optimized for new infection identification.

Main Methods:

• Development of a privacy-preserving DFL-based LeNet model for infection identification.
• Implementation of automatic effective feature selection tailored for new infection detection.
• Evaluation of the model's performance using real-world genome sequence data distributed across multiple clients.

Main Results:

• The proposed DFL model achieved an overall accuracy of 99.12%.
• Performance metrics include 98.23% precision, 98.04% recall, and 96.24% F1-score.
• The model demonstrated significant improvements over benchmark models, with ROC AUC at 98.24% and Cohen's kappa at 83.94%.

Conclusions:

• The developed DFL model offers a secure and accurate alternative for identifying new infections from GSs.
• The model effectively balances high performance with stringent patient data privacy and security requirements.
• Empirical results support the model's potential for broader application in identifying diverse virus strains.