Developing a Federated Machine Learning Framework for Antimicrobial Resistance Prediction from Genomic Data Across Nigerian Tertiary Hospitals Without Data Sharing

Accurate prediction of antimicrobial resistance (AMR) phenotypes from bacterial whole-genome sequences has the potential to transform clinical decision-making speed and precision. However, applying machine learning to AMR prediction at national scale in Nigeria is impeded by data fragmentation, patient privacy regulations, and limited central genomic infrastructure. Federated learning, which enables model training across distributed datasets without transferring raw genomic data, offers a theoretically and practically compelling solution. This dissertation develops a federated machine learning framework for WGS-based AMR phenotype prediction deployable across geographically distributed Nigerian tertiary hospitals. Institutional genomic and corresponding phenotypic AMR data were contributed from six tertiary hospitals spanning all geopolitical zones: Aminu Kano Teaching Hospital (Kano), Lagos University Teaching Hospital (Lagos), University of Nigeria Teaching Hospital (Enugu), University of Calabar Teaching Hospital (Calabar), University of Ilorin Teaching Hospital (Ilorin), and Federal Medical Centre Birnin Kebbi (Kebbi). A total of 2,840 bacterial genomes with paired broth microdilution MIC data were used. A PySyft-based federated learning pipeline was developed and validated, comparing federated model performance against centralized and local model baselines using ROC-AUC and F1 metrics. Separate federated models were trained for K. pneumoniae (n=1,142), E. coli (n=876), and S. aureus (n=822) targeting key resistance phenotypes. Federated models achieved ROC-AUC values within 3.2% of the centralized training benchmark for major resistance categories, including beta-lactams (AUC 0.94) and fluoroquinolones (AUC 0.91). Communication-efficient gradient compression reduced inter-site data transmission by 94% without model performance degradation. The framework constitutes an original machine learning and molecular biology infrastructure contribution to Nigeria's AMR surveillance ecosystem. Keywords: federated learning, AMR prediction, whole-genome sequencing, machine learning, Nigeria.