Authors: Deepa Shree R
Abstract: The rapid rise of antimicrobial resistance necessitates robust diagnostic tools that can integrate heterogeneous clinical data across decentralized healthcare institutions while ensuring patient data confidentiality. To address these requirements, this framework leverages federated learning to enable collaborative model training across institutions without necessitating data centralization (Zwiers et al., 2024), while simultaneously incorporating Bayesian inference to quantify predictive uncertainty and enhance clinical interpretability (Kumari et al., 2026). Furthermore, by synthesizing multimodal electronic health records, the proposed architecture overcomes the limitations of centralized data silos and addresses the inherent challenges of data bias and clinical validation in AMR research (Hardan et al., 2024; Narra et al., 2024). Antimicrobial resistance makes infections harder to treat, so we need faster and more reliable ways to detect resistant pathogens. Artificial intelligence can help predict resistance using genomic and other biological data (Lastra et al., 2024). However, many current methods depend on centralized datasets (Zwiers et al., 2024), offer little insight into how certain their predictions are, and are difficult to interpret in high-stakes settings (Kumari et al., 2026). In addition, healthcare and biological data are often spread across institutions and cannot be freely shared because of privacy, governance, and regulatory requirements. This paper presents FedBioGuard, a privacy-preserving framework for predicting antimicrobial resistance. It combines federated learning, multimodal data, uncertainty estimates, explainable AI, and evidence-based large language model support. The main goal is to examine whether this approach can make reliable AMR predictions across institutions with different types of data, without sharing raw patient records. Each participating institution keeps its data locally and helps train a shared model by sending model updates. The model can use genomic data, microbiome or metagenomic data when available, and structured clinical information. An uncertainty module shows how reliable each prediction may be, while explainability methods highlight the biological and clinical features that influence the results. A retrieval-augmented language model is used only after prediction to turn the results into clear summaries supported by scientific evidence. By decentralizing the training process, FedBioGuard mitigates the risks associated with data privacy while addressing the limitations of traditional cultivation-based diagnostics that are often too slow to guide immediate treatment decisions (Dayan et al., 2021; Inda-Díaz et al., 2026). By addressing technical hurdles such as data heterogeneity and the "black box" nature of deep learning, this framework fosters the development of transparent, robust clinical decision support systems (Cavallaro et al., 2023). The evaluation will compare centralized, local-only, single-modality, and federated models using simulated data from multiple institutions. The models will be assessed for predictive performance, calibration, uncertainty quality, performance under distribution shifts, and consistency of their explanations. The study aims to provide a reproducible way to examine how privacy-preserving collaboration, multimodal learning, and uncertainty estimation can work together for AMR prediction. FedBioGuard is not intended to replace laboratory susceptibility testing; instead, it is a research framework for studying trustworthy AI-assisted AMR analysis.
