Antimicrobial resistance (AMR) is rapidly becoming a major global health threat with the World Bank estimating that AMR could result in US$ 1 trillion additional healthcare costs by 2050, and US$ 1–3.4 trillion gross domestic product losses per year by 2030. It has been reported that poor access to quality diagnostics is one of the main contributing factors, among others (1). The advent of genomic sequencing has significantly enhanced our ability to understand the genetic mechanisms behind resistance, track their evolution over time, and accurately monitor the spread of resistant strains.

However, despite its analytical power, the integration of genome sequencing into routine clinical workflows remains limited. Barriers such as cost, personnel training requirements, and longer turnaround time (TAT) compared to traditional phenotypic testing or polymerase chain reaction (PCR) assays continue to delay its broader adoption (2).

In parallel, artificial intelligence (AI) has been proposed as a revolutionizing tool for laboratory sample processing and data management. Specifically, machine learning (ML) systems, with the ability to learn and improve from training datasets, are emerging with a central role in clinical microbiology (3). Beyond ML, more sophisticated AI frameworks, including convolutional neural networks (CNNs) and large language models (LLMs) are being investigated for their ability to handle complex, high-dimensional datasets and generate clinically relevant outputs (4–6).

Recent studies suggest that AI may play an integrative role across the full continuum of pathogen genomics, from raw data processing to the generation of actionable insights, potentially addressing several obstacles that limit the scalability of sequencing-based diagnostics (7). Here, we examine how AI-driven approaches can enhance genomic surveillance, resistance prediction, and the broader fight against AMR when strategically combined with next-generation sequencing (NGS) and complementary technologies.

This narrative review was conducted to synthesize and critically appraise current evidence on the integration of AI with genomic and rapid diagnostic technologies for AMR detection and surveillance in hospital and public health settings. Abbreviations, and the corresponding brief definitions, of all AI/ML-related terms used in this manuscript are provided in Table 1. The review was informed by a predefined conceptual PI(C)O framework (Supplementary Table S1), which served as a scoping scaffold to define the population, intervention domains, and outcomes of interest, rather than as a formal systematic or scoping review protocol. The literature search was conducted across three major biomedical databases: MEDLINE (via Ovid), Embase, and the Cochrane Central Register of Controlled Trials (CENTRAL). Searches included publications up to April 2025, with particular emphasis on studies published within the last decade, reflecting the rapid evolution of AI methodologies and sequencing technologies. Search concepts combined terms related to AI and ML (e.g., “AI,” “machine learning,” “deep learning,” “natural language processing”), genomic and rapid diagnostic technologies (e.g., “next-generation sequencing,” “whole-genome sequencing,” “metagenomics,” “nanopore sequencing,” “MALDI-TOF MS,” “Raman spectroscopy”), AMR, and surveillance or infection prevention workflows.

The scope of this review is intentionally restricted to AI applications integrated with genomic and rapid diagnostic technologies for AMR detection, characterization, and surveillance. It does not aim to provide a comprehensive review of AI tools focused primarily on patient-level MDRO carrier prediction, hospital logistics, or infection control management, which fall outside the focus of the present manuscript.

The review prioritized studies relevant to human health, particularly those addressing hospital or inpatient settings, population-level surveillance, infection prevention and control (IPC), and antimicrobial stewardship. One Health or environmental studies were included when they directly informed surveillance frameworks, transmission dynamics, or public health implementation. We explicitly excluded studies limited to: (i) AI applications focused solely on antimicrobial prescribing optimization or laboratory automation without infection control relevance, and (ii) theoretical models, simulations, or proof-of-concept frameworks lacking a clear pathway to real-world application.

Priority was given to multicenter studies, externally validated models, real-world or near-real-world implementations, and consensus documents or guidelines where available. Single-center and early proof-of-concept studies were included selectively to illustrate emerging approaches, but were interpreted cautiously with respect to generalizability, validation, and readiness for implementation. Up-to-date narrative reviews were included to enhance the completeness of the literature coverage. During manuscript revision, nine relevant articles published after April 2025 were identified and incorporated as substantive updates in the field, and four studies initially excluded were subsequently included following re-evaluation of their relevance.

This review is intentionally narrative and non-exhaustive. No formal systematic review methodology, risk-of-bias assessment, or quantitative synthesis was undertaken. The objective was to provide a structured and critical overview of key AI-enabled genomic and rapid diagnostic applications for AMR surveillance, highlighting methodological limitations, operational constraints, and implications for public health practice.

The critical threat of AMR is prompting the need for advanced computational methods capable of rapid and precise resistance prediction. Recent developments in ML leverage genomic data, pan-genome analyses, and ensemble techniques to enhance detection accuracy. Nevertheless, challenges persist in feature selection, model generalizability, and validation across heterogeneous datasets.

A foundational step in AMR prediction is identifying informative genomic features. Traditional models often rely on known resistance genes or mutations, which may overlook novel or combinatorial signals. Her and Wu demonstrated that pan-genome-based feature extraction improved prediction accuracy in Escherichia coli by capturing accessory genome elements linked to strain-specific resistance mechanisms (8). Notably, non-core genes, present only in subsets of strains, outperformed core-genome features, even though not all directly conferred resistance. Many of these genes, in fact, often carried on mobile elements like plasmids, contributed to broader multidrug resistance signals. The proposed AI-based genetic algorithm approach based on non-core genes identified informative clusters that discriminated resistant from susceptible E. coli, achieving area under the curve (AUCs) above 0.9 for ampicillin, gentamicin, trimethoprim-sulfamethoxazole, and ciprofloxacin, surpassing conventionally examined clusters. Notably, the model flagged the pmrC/pmrE cluster (classically associated with polymyxin resistance) as predictive not only of polymyxin resistance but also of ampicillin and trimethoprim–sulfamethoxazole resistance, underscoring the capacity of ML to reveal latent resistance patterns.

Similarly, Wang et al. (9) showed that combining ML with whole-genome sequencing (WGS) predicted Staphylococcus aureus minimum inhibitory concentrations with >95% accuracy, particularly excelling in detecting cefoxitin resistance and reliably identifying methicillin-resistant S. aureus (MRSA) strains.

In this context Hyun et al. (10) developed a pan-genome framework for S. aureus, P. aeruginosa, and Escherichia coli, creating a binary genome-by-feature matrix to train multiple support vector machines (SVM) models. Their ensemble approach, based on randomly sampled strain-feature subsets, achieved accuracies between 79.3 and 99.5%, outperforming classical statistical and ranking methods. Impressively, the model recovered 45 known resistance genes and suggested 25 novel candidates (10). By training every sub model on a randomly selected part of the dataset, they reduced overfitting and improved model generalizability.

Álvarez et al. (11), for instance, applied SVM to 5,299 A. baumannii genomes, including 512 global clone 1 (GC1) and 4,787 non-GC1, and identified a previously unknown 367 base pair (bp) fragment of moaCB, termed U1, uniquely defining the high-risk GC1. The model achieved notable performance (sensitivity 1.00, specificity 1.00, accuracy 1.00; 0 false positives and 0 false negatives), illustrating how ML can reveal clinically actionable biomarkers and directly support rapid diagnostic development in outbreak scenarios (11).

Most studies validate on single datasets, risking inflated performance estimates. When this is not the case, model performance is not consistent across different settings, and often fails to replicate across hospitals, likely due to differences in sequencing platforms, population structure, and metadata quality. On the other hand, while multicenter datasets enhance robustness, they also introduce new biases from heterogeneous lab protocols. Federated learning presents a promising alternative, allowing multi-institutional training without sharing raw genomic data. This approach can balance generalizability with privacy, laying the foundation for regional ML-based AMR early-warning systems (12).

Future models may benefit from integrating multi-omics layers, such as transcriptomics or proteomics, to uncover complex interactions or cryptic features that lead to AMR.

Thanks to rapid advancements in sequencing technologies, metagenomics has the potential to provide a detailed view of the entire microbial ecosystem within clinical and environmental samples. Metagenomics provides a panoramic, high-resolution snapshot of the entire microbial genomic landscape, representing a substantial advance in microbiological diagnostics by moving beyond the targeted detection of individual pathogens and capturing comprehensive genomic information even in the absence of culture growth.

For example, Fida et al. (13) observed that 16S rRNA gene metagenomics was able to identify a potentially pathogenic organism in 47% of septic patients compared to 32% with blood cultures. As sequencing technologies continue to advance, this approach may potentially bypass traditional methods in the future, beginning with direct nucleic acid extraction from the collected specimen.

These technologies, like shotgun metagenomic sequencing, generate massive volumes of raw genetic data. Therefore, scaling this approach requires sequencing and data analysis technologies capable of managing the complexity and volume of information generated (14).

Hence, AI-based data management could be of paramount importance since the task of assembling millions of short reads into a coherent biological narrative is expected to be very time consuming without advanced computational pipelines. AI algorithms are used not only to align reads with reference databases, but more importantly, to perform deeper reconstruction and inference tasks. ML algorithms, specifically deep learning (DL) methods like CNN, could address several aspects of metagenomics sequencing, including microbial population analysis, novel pathogen detection, sequence classification, patient stratification and disease prediction (15).

Additionally, they have been successfully applied to functional annotation, generating comprehensive taxonomic profiles, classifying whether a given sequence is an antibiotic resistance gene, identifying all resistance determinants that together compose the sample’s resistome, and detecting virulence genes or metabolic biomarkers (16, 17).

This broader, systems-level, insight can redefine how antimicrobial resistance is understood and monitored. While conventional methods profile resistance at the level of individual isolates, metagenomics unveils the resistome embedded within a sample’s microbiota, including silent or subdominant reservoirs of resistance that may be mobilized under antibiotic pressure (18). This capacity is especially valuable for colonization screening, where characterizing the full spectrum of resistance genes enables more informed infection control and outbreak prevention strategies.

In a 2025 study, Mikolas et al. (19), through shotgun sequencing on Illumina platform, applied high-throughput metagenomic analysis to 96 intensive care unit (ICU) patient, staff, and environmental samples, aiming to identify early microbial biomarkers for prognostic assessment and uncovered extensive AMR reservoirs. Their computational models identified early oropharyngeal resistome signatures that discriminated early- from late-mortality patients with strong prognostic performance (AUC up to 0.90), illustrating how advanced metagenomic analytics can extract clinically actionable signals from complex datasets to support patient-centered clinical management strategies in settings where rapid clinical deterioration is frequent (19).

A promising application of AI to metagenomics is represented by the ability to differentiate between viable and dead microorganisms, addressing a common challenge in modern microbiology where molecular positivity cannot distinguish live from dead organisms. In standardized laboratory conditions, Ürel et al. (20) developed a model capable of detecting the viability of bacteria with an accuracy of 0.83 and an AUC value of 0.90 through the analysis of shotgun nanopore sequencing data.

Yet, the promise of this panoramic vision must be tempered by current limitations. The perceived high costs of sequencing, coupled with the need for robust bioinformatic infrastructure and expertise, currently restrict metagenomics to high-stakes clinical scenarios, such as culture-negative infections (e.g., endocarditis), intraoperative diagnostics, or precision outbreak tracking.

However, it should be noted that in a 2021 cost-effectiveness simulation model of a 32-month carbapenem-resistant A. baumannii (CRAB) outbreak, early use of WGS and shotgun metagenomics was predicted to avert 18 additional infections, yield 74 extra quality-adjusted life years (QALYs), and reduce overall hospital expenditures by AU$93,822 compared with the real-life outbreak management scenario, highlighting the potential clinical and economic value of sequencing-guided infection control (21). Then, the integration into routine, high-throughput clinical microbiology should move from aspirational to practical.

During and after SARS-CoV-2-caused COVID-19 pandemic, pathogen WGS shifted from reduced capacity to surveillance backbone, enabling rapid lineage tracking, transmission mapping, and detection of emerging resistance across clinical and public-health settings. Consensus guidance now frames genomic surveillance as essential for infection control and AMR monitoring across One-Health domains (22).

The term “real-time” should not be interpreted as “instantaneous.” It should refer to a data flow that makes information timely for effective infection control decisions and public health policies. Real-time genomic surveillance should refer to the continuous process of monitoring and analyzing the genetic material of pathogens and AMR-associated distribution, as they circulate in a population, with the goal of providing actionable insights to control outbreaks as soon as possible as they occur.

Unlike traditional surveillance, which often relies on retrospective data, real-time surveillance should aim for a fast “sample-to-result” TAT, typically within days or weeks, to allow public health officials to intervene during an active transmission event rather than simply documenting it after the event.

Real-time genomic surveillance has been proposed to anticipate the emergence of resistant strains, track their spread, link human, animal, and environmental data, and support rapid response.

As partially anticipated, currently, two complementary sequencing paradigms are fundamental for surveillance: (i) isolate WGS, which delivers high-resolution phylogenetics and precise AMR determinant calling, and (ii) metagenomic NGS (mNGS), which profiles mixed specimens (e.g., blood, cerebrospinal fluid, bronchoalveolar lavage, wastewater) without culture.

A 7-year clinical mNGS evaluation highlighted stable diagnostic yield but also the operational drivers of TAT: 8.2–11.4 days from sample collection to result and 3.6–3.8 days from start of processing to result. It emphasizes the need for workflow engineering and automation that include the pre-analytical phase (decision to order testing, sample shipping, batch testing) to achieve true “real-time” (23). AI-assisted triage could be a practical target for improvement of the pre-analytics. Examples of AI-based systems applications include auto-eligibility of cases that might benefit from NGS (e.g., culture-negative samples), automated creation of samples shipping orders, and efficient batching of a sufficient number of specimens for NGS (24).

In 2021, in a study on new workflow for real-time prediction of pathogenic microorganisms, Bartoszewicz et al. (25) built DL models specialized in inference from incomplete short- and long-read sequencing data. The models showed an accuracy higher than 78.5% in inferring whether an Illumina or a Nanopore sequence belongs to pathogenic bacteria or viruses, for reads as short as 50 bp obtained during the Illumina sequencing run or after the first 250 bp (0.5 s of sequencing time) of Nanopore reads (25). If properly validated in the clinical context, such AI-based systems, performed in real-time during the sequencing run, could potentially save hours (depending on the platform used) in the post-analytical phase of pathogens identification.

However, Liu et al. (26) showed that adaptive nanopore sequencing, even without AI capabilities, already enabled reliable pathogen identification using the reads obtained within the first hour of sequencing, with a categorical agreement (CA) of 89.3%. Overall, it achieved 90.6% CA for AMR prediction with very major error rates of only 1–3% in key pathogens, and detected 23 polymicrobial infections, including 7 missed by standard diagnostics, substantially improving timely stewardship and outbreak control (26).

Beyond diagnostics, long read-based technologies like nanopore resolve complex AMR loci and structural variants that short-read pipelines may miss, strengthening surveillance signals about transmission of resistance cassettes (2).

Multiple evaluations now show that embedding WGS into infection-prevention programs improves timeliness and precision of outbreak recognition, with measurable patient and economic benefits when results are integrated into routine workflows. Optimized pipelines can move from sample to action in a clinically relevant window, particularly when coupled to automated analytics and visualization for infection prevention and control teams (27). This combination of AI and WGS has been successfully evaluated for outbreak investigation in hospital settings. Sundermann et al. (28) developed an AI algorithm that helped identify likely transmission routes after genomic clustering was established. The model increased the identification of epidemiological links by 3.8% as compared to traditional review methods of infection prevention and control, with an overall sensitivity of 91.7% (28).

For population-scale surveillance studies, including resistome surveillance, metagenomic analysis of sewage has been proposed as a powerful tool, capturing antibiotic resistance gene (ARG) diversity and trends independent of testing behavior. Large international datasets demonstrated global heterogeneity of ARGs and their association with socio-economic and environmental factors, while newer studies use serial wastewater metagenomes to relate resistome shifts to antimicrobial use and interventions, positioning wastewater-based epidemiology (WBE) as a sensitive barometer for public-health AMR policies (29, 30). WBE might be integrated with AI-based tools to improve surveillance capabilities. Liu et al. (31) demonstrate that ML models trained on WGS data using gene presence and absence profiles, SNP variants, and k-mer features can reliably predict antimicrobial resistance in S. aureus. In a dataset of 1,764 isolates tested against ten antibiotics, their best performing models achieved AUC values ranging from 0.83 to 0.99, with six antibiotics reaching an AUC above 0.99 and nine exceeding 0.96. These findings show that rich genomic feature sets enable ML classifiers to capture resistance determinants with consistent accuracy across various antimicrobial agents (31).

Emerging studies clarify that “real-time” also involves an efficient integration of sample logistics, run time, computation, expert review, and feedback into decisions. Reported clinical mNGS experiences indicate that shaving days off logistics and report generation often dwarfs gains from faster chemistry alone, areas where AI-driven automation (eligibility triage, templated narratives, anomaly alerts) can reduce wall-clock time (32).

While this review centers on genome-based AI for AMR prediction, emerging non-genomic tools like Raman spectroscopy and matrix-assisted laser desorption/ionization-time of flight mass spectrometry (MALDI-TOF MS) can complement these approaches. When paired with ML, they offer rapid, label-free resistance profiling, especially useful when sequencing is unavailable or delayed.

The methodologies reviewed in this paper, specifically the application of AI to genomic strategies, offer transformative opportunities for pathogen identification and AMR detection. These AI-driven tools hold the potential to substantially reduce diagnostic TAT, enabling clinicians to initiate targeted therapies with greater promptness and appropriateness. As a result, the implications of AI-aided sequencing and spectroscopy for public health are profound: at the bedside, ML models applied to MALDI-TOF or Raman spectra can deliver rapid AMR predictions, offering valuable support for early treatment decisions and antimicrobial stewardship. Further upstream, WGS and mNGS provide higher-resolution insights, enabling the detection of transmission clusters, characterization of resistance mechanisms, and tracking of high-risk clones across healthcare facilities and regions. These genomic approaches may also underpin sentinel surveillance programs and broader population-level initiatives, such as national strain monitoring and prevalence mapping, though they require robust infrastructure, standardized workflows, and centralized coordination. Wastewater-based epidemiology is emerging as a complementary strategy, leveraging mNGS to detect shifts in community-level resistomes and early signs of emerging threats, though analytical complexity remains a barrier to widespread implementation. In low-resource settings, simplified and portable solutions, such as real-time nanopore sequencing or MALDI-based species identification, can provide actionable insights with lower technical requirements, provided there is investment in training and system integration.

The near future likely lies in a tiered diagnostic framework, which may help balance speed, cost, and resolution: for instance, deploying MALDI-TOF or Raman-based ML for rapid, priority triage, while reserving WGS or mNGS for high-risk outbreak scenarios, such as MDRO outbreaks in large ICUs or oncohematological facilities, that require deeper genomic resolution and contextual interpretation. Moving forward, the integration of population-based metrics other than hospital systems’ EHRs may warrant future integration of AI in public health interventions against AMR such as improving vaccine uptake in fragile populations or assist and monitor the prescription of antimicrobials in primary care (44, 45).

Such a stratified approach balances diagnostic depth with operational feasibility, ensuring that high-resolution insight is applied precisely where it is needed most.

Taken together, the evidence reviewed indicates that AI has the potential to substantially enhance clinical microbiology and AMR surveillance, but its impact will depend on disciplined translation rather than technological novelty alone. Progress will require robust validation, transparent reporting of clinically meaningful errors, and seamless integration of AI outputs into resource-aware and coherent diagnostic workflows. Only through such a rigorous and pragmatic path can AI move from isolated proofs of concept to reliable components of routine microbiological practice.