Federated learning framework for predicting multi-drug resistant tuberculosis across regional databases.

Area of Science:

• Computational biology and bioinformatics
• Machine learning in healthcare
• Epidemiology and public health

Background:

• Multi-Drug Resistant Tuberculosis (MDR-TB) poses a significant global health challenge, particularly in resource-limited settings.
• Traditional diagnostic methods for MDR-TB are time-consuming, expensive, and inaccessible in many regions.
• Existing machine learning approaches for MDR-TB prediction face data privacy concerns and regulatory hurdles for data sharing among healthcare institutions.

Purpose of the Study:

• To develop a privacy-preserving federated learning framework for predicting MDR-TB.
• To enable collaborative model training across multiple healthcare organizations without direct data sharing.
• To assess the performance of the federated learning model against centralized approaches.

Main Methods:

• A federated learning framework was designed to facilitate collaborative model building using genomic and clinical data from diverse healthcare centers.
• The framework allows multiple institutions to train a shared MDR-TB prediction model without pooling sensitive patient information.
• The model's performance was evaluated considering data heterogeneity across local datasets.

Main Results:

• The federated learning model achieved high predictive performance, with an Area Under the Curve (AUC) close to 91% and accuracy around 89%.
• Real-world testing on a dataset including whole-genome sequencing and clinical data demonstrated accuracy above 88% and precision/recall above 86%.
• The model exhibited consistent results across different local datasets, indicating robustness and generalizability.

Conclusions:

• Federated learning offers a secure, adaptable, and accurate solution for MDR-TB prediction across distributed databases.
• This approach effectively balances the need for data utilization with stringent privacy requirements, promoting wider adoption in disease management.
• Future research will focus on incorporating diverse data types and enhancing data transmission security for improved clinical utility.