Perhaps the most fundamental obstacle to integrated AMR surveillance is the legal, ethical and competitive constraints that prevent the sharing of sensitive data across institutional and sectoral boundaries [25]. Federated learning directly addresses this challenge by enabling AI model training across distributed datasets without requiring raw data movement, thereby substantially reducing, though not eliminating, as discussed below the privacy burden while enabling collaborative intelligence [26].

A cross-sectoral (human-animal-environmental) federated learning deployment specifically for AMR surveillance has not, to our knowledge, yet been documented in the peer-reviewed literature; readers should treat this as an unrealized but technically feasible opportunity rather than an established practice. Proof-of-concept evidence from clinical federated learning is nonetheless instructive: in a ten-institution consortium training a shared model without any raw patient data leaving participating sites, the resulting model reached 99% of the predictive quality achieved by a fully centralized model trained on pooled data [27]. Translating this architecture to link hospital, veterinary and environmental surveillance nodes for AMR would preserve data sovereignty, accommodate diverse regulatory frameworks and enable cross-sectoral collaboration that centralized data-sharing arrangements currently preclude provided the safeguards discussed below are in place.

This last caveat matters. Federated learning reduces, rather than eliminates, disclosure risk: shared model updates and gradients can still be vulnerable to gradient-inversion or membership-inference attacks capable of partially reconstructing source data [25]. Treating federation alone as a sufficient privacy guarantee would therefore be premature. Realizing its privacy benefits in practice requires complementary safeguards secure aggregation protocols that prevent any single party from inspecting individual updates, formal privacy-risk assessments conducted before deployment and clear institutional accountability for model governance across participating sites [26].

For low- and middle-income countries (LMICs), where data protection infrastructure may be nascent and data governance frameworks often underdeveloped, federated learning offers a technically feasible pathway to participation in global surveillance networks without compromising local data sovereignty [28]. This is particularly significant given that LMICs bear the greatest AMR burden yet often face the most stringent constraints on data sharing [29].

The data that could inform AMR surveillance exist in an extraordinary variety of forms: structured laboratory results, free-text clinical narratives, veterinary diagnostic reports, environmental monitoring records, agricultural production data and policy documents [30]. Natural language processing (NLP) offers a powerful toolkit for extracting structured, machine-readable information from these heterogeneous sources, thereby enabling semantic interoperability the capacity to understand and integrate information across diverse data representations [31].

Recent advances in transformer-based language models have dramatically expanded NLP’s capabilities in the biomedical domain [32]. Researchers have successfully employed machine learning frameworks combining transformer architectures and cluster analysis to analyze over 44,000 scientific publications, identifying key AMR research topics, subtopics and detection methods across One Health sectors [33]. More ambitiously, large language models have been developed to assist AMR policy development by synthesizing evidence from national action plans across 146 countries [34]. The AMR-Policy GPT, a large language model with advanced retrieval-augmented generation capabilities, can search and summarize evidence from plans, metadata and technical knowledge, providing traceable references from global document databases [34].

It is important to separate what this evidence actually demonstrates. The topic-clustering application [33] and the policy-synthesis application (AMR-Policy GPT) [34] both show NLP performing well on text corpora scientific literature and policy documents, respectively that are relatively well-structured and written in standardized professional language. Neither constitutes evidence that NLP can reliably extract and harmonize information from the routine, often noisy clinical, veterinary and environmental surveillance records that this paper’s interoperability argument actually depends on. Peer-reviewed evaluations of NLP performance specifically on cross-sector routine-record harmonization remain limited; the methodological and ontological groundwork exists [30][31], but validated performance evidence at the scale needed for operational surveillance is, at present, an evidence gap rather than a demonstrated capability.

The implications of these developments are profound. By converting unstructured text into standardized, machine-readable data, NLP enables the data integration essential for true One Health surveillance. However, critical challenges remain: NLP models must be trained on representative data, evaluated for bias and continuously validated to ensure accuracy across diverse linguistic and cultural contexts [35].

Even the most accurate predictive models will fail to influence practice if clinicians, veterinarians and policymakers cannot understand or trust their outputs [36]. The escalating challenge of AMR demands clinician-ready machine learning models that are not merely accurate but interpretable [37]. By treating resistance genes as independent features and augmenting them with curated single-nucleotide polymorphisms and contextual markers, explainable AI approaches deliver scalable, transparent predictions aligned with clinical decision-making needs [38].

Transparency in AI predictions empower evidence-based prescribing decisions, reduces treatment failure and enables tailored care for patient subgroups at greatest risk of resistance [39]. When models remain interpretable, predictions align with the clinician’s mental model of resistance mechanisms, thereby bridging the gap between algorithmic output and clinical action [40]. In the high-stakes environment of AMR management, where treatment decisions have life-or-death consequences, explainable AI is not merely a technical preference but an ethical imperative [41].

Table 1 summarizes these three modalities, the interoperability dimension each addresses and their respective benefits and limitations.

Table 1.Comparison of AI techniques for interoperable One Health AMR surveillance.

AI models are only as robust as the data on which they are trained [43]. Yet across human, animal and environmental health sectors, data are collected using disparate methods, definitions and formats that resist harmonization. Divergent analytical methodologies such as culture-based versus genomic approaches generate data that cannot be meaningfully compared, partially explaining why One Health solutions have been slow to materialize [7].

The challenge extends beyond technical data formats to encompass fundamental ontological differences: how different sectors conceptualize and classify resistance, disease and health. For instance, the human health sector typically classifies resistance by clinical significance for patient outcomes, while the veterinary sector emphasizes implications for food production and animal welfare and the environmental sector focuses on ecological persistence and transmission dynamics [44]. These different conceptual frameworks generate data that are not merely stored differently but ask fundamentally different questions [45].

Greater harmonization of research methods, data standards and ontological frameworks across sectors is urgently needed [7]. Effective One Health information systems depend on standardized data protocols, interoperable platforms and strong data sharing networks [46]. Without common standards, ontologies and application programming interfaces, even the most sophisticated AI will struggle to integrate information meaningfully [47].

In LMICs, where the AMR burden is often greatest, limited digital access, funding shortages and lack of technical expertise severely constrain AI implementation [48]. Challenges related to user engagement and incentives, data quality and standardization, infrastructure and connectivity, capacity building and policy alignment are consistently identified as barriers to effective digital One Health surveillance [49]. This unevenness in equipment and technical capacity across countries particularly acute in the low- and middle-income settings where antimicrobial use and resistance are often highest means that any AI-enabled surveillance strategy must be phased and context-appropriate rather than assuming uniform readiness for deployment.

Even in countries with robust infrastructure, incompatible databases, siloed institutions and lack of common standards impede progress [50]. The challenges are compounded by the need for sustained capacity building, reliable electricity and internet connectivity that many regions lack [51]. Investment in capacity development in LMICs must be prioritized, as these are settings where the AMR burden is most acute and the benefits of effective surveillance greatest [49].

Data sharing across sectors raises complex questions of ownership, usage rights and privacy [52]. Clear governance frameworks defining these parameters are essential for confident AI deployment, yet such frameworks are often absent or inconsistent across jurisdictions [53]. The Quadripartite organizations have called for coordinated surveillance systems and launched a Joint Offer for targeted One Health investment, but translating this vision into operational reality requires national-level political commitment and investment that remains uneven [54].

The fragmentation of regulatory frameworks creates particular challenges for federated learning approaches, which operate across institutional and jurisdictional boundaries. Without clear guidance on data ownership, accountability and liability, institutions remain reluctant to participate in cross-sectoral AI initiatives [55].

Investment in AI for AMR must prioritize foundational interoperability infrastructure over isolated predictive tools [57]. This means dedicating resources to common data standards, shared platforms, governance frameworks and capacity building—not merely algorithmic innovation [58]. When embedded within well-regulated One Health frameworks, AI can strengthen surveillance, improve treatment decisions and support evidence-based interventions [59]. However, this requires deliberate investment in the foundational architectures that enable data to flow across sectors [60].

Explainability must be treated as a non-negotiable requirement for clinical AI deployment [61]. Transparency empowers evidence-based prescribing decisions and reduces treatment failure [38]. AI should be positioned as a decision-support tool that augments, rather than supplants, clinical reasoning [62]. By keeping models interpretable, predictions align with clinical mental models of resistance mechanisms, enabling trust and adoption [40]. This is not merely a technical preference but an ethical and practical imperative for effective AMR management [41].

Solutions must be context-appropriate rather than imposing high-resource models on LMIC settings [63]. Offline-capable mobile applications, low-bandwidth platforms and locally relevant training programs may be more suitable than sophisticated cloud-based systems [64]. Co-design with local stakeholders, infrastructure investment, applications integration, cross-sectoral collaboration, continuous training, evidence-based policy support and sustainable funding models are crucial for building effective, inclusive and scalable ICT-driven solutions [65].

The current political momentum provides an unprecedented opportunity [66]. The Quadripartite Joint Offer for targeted One Health investment and the inclusion of AMR in the WHO Pandemic Agreement negotiations signal growing political will [67]. However, opportunity alone is insufficient. Realizing AI’s potential to combat AMR requires deliberate, coordinated and adequately funded effort to build the interoperable information systems that One Health demands [67].

The priorities above remain aspirational unless translated into an actionable sequence. We propose the following operational steps, each a precondition for the next: 1) agree a minimum set of shared data elements that every participating sector commits to reporting in a common format; 2) map these elements onto existing standards and ontologies, resolving the definitional discordance identified in Section 4.1 before any data are pooled or federated; 3) validate data quality at each participating site against agreed thresholds, since AI trained on unvalidated data will simply propagate existing inconsistencies; 4) secure governance and ethics approval covering data ownership, accountability and liability across all participating institutions and jurisdictions; 5) pilot a federated deployment at limited scale, incorporating the privacy safeguards discussed in Section 3.1; and 6) scale deployment only after pilot results are independently evaluated, with continuous monitoring for model drift and equity of access. Sequencing matters: attempting model deployment before data-quality validation and governance approval are in place has been a recurring failure mode in digital One Health initiatives [49][50].