Uncertainty-Aware Antimicrobial Resistance Prediction in E. coli and S. aureus Isolates Using Hybrid Bayesian Neural Networks

Background: Antimicrobial resistance (AMR) poses a critical global health threat. However, conventional detection methods still require up to 72 hours, leading to treatment delays. AI models trained on mass spectrometry data enable faster prediction, but current approaches often lack uncertainty estimation, cross-hospital generalization, and clear interpretability of their predictions. This study proposes a framework for AMR prediction that quantifies uncertainty while enhancing interpretability and generalizability. Methods: We developed species-specific AI models for AMR prediction and incorporated Bayesian inference to estimate predictive uncertainty. Models were trained on the largest DRIAMS subset, corresponding to data from a single hospital. Generalization was evaluated through fine-tuning and zero-shot testing on the remaining three DRIAMS subsets, each originating from a different hospital, as well as on the MS-UMG dataset from a different country. Model interpretability was examined using SHAP-based feature attributions and UMAP visualizations. Results: Proposed AI models outperformed state-of-the-art methods. On held-out testing from the source DRIAMS dataset, the models achieved AUROC and AUPRC values of up to 0.90, with balanced accuracy reaching up to 0.80. Performance remained strong under distribution shift, attaining AUROC and AUPRC values of up to 0.88 and balanced accuracy up to 0.78 following fine-tuning on data from other hospitals, and up to 0.81 AUROC/AUPRC with balanced accuracy up to 0.71 in zero-shot evaluation. Bayesian AI models matched this performance while providing meaningful uncertainty estimates, which were found to be higher for misclassified samples. SHAP and UMAP analyses revealed biologically meaningful representations, with UMAP showing clear separation across most antibiotics and SHAP indicating consistent feature importance within same antibiotic classes for each species. Conclusion: This study demonstrates that proposed AI models generalize well across different datasets. Elevated uncertainty for misclassifications indicates that the models appropriately flagged ambiguous cases. These findings suggest that our approach represents an important step towards building more robust and trustworthy AMR prediction systems for clinical and surveillance applications.