Role of artificial intelligence in bacterial diagnostics and surveillance of anti-microbial resistance

Abstract

Antimicrobial resistance (AMR) is a serious global health threat caused by the overuse and misuse of antibiotic resulting in treatment failure. The current conventional techniques have various constraints requiring specialized expertise, longer turnaround times requiring rapid point of care and transformative solutions. This narrative review explores the applications of artificial intelligence (AI) in AMR diagnostics. A structured search of PubMed, Web of Science, and Google Scholar was conducted using the MeSH terms. Relevant studies were screened and synthesized in four themes; phenotypic and genotypic identification, antimicrobial susceptibility testing (AST), AMR surveillance, and antibiotic development. The reporting was guided by the Scale for the Assessment of Narrative Review Articles (SANRA). Across AMR diagnostic, machine learning and deep learning improved the accuracy, reproducibility, and scalability of bacterial identification by learning complex patterns of AMR. The AI models utilized wide variety of data including genomic profiles, radiological imaging, microscopy, agar plate photography, and biochemical signatures such as MALDI TOF mass spectrometry. In Antibiotic Susceptibility Tests (AST), AI helped in standardizing the interpretation of disc diffusion and MIC assays. In AMR surveillance, AI models supported screening and genomic detection of resistance determinants, which enabled the identification of resistance trends and policy evaluations of One Health integration. In antibiotic development, AI contributed to therapeutic discoveries through screening large chemical libraries and designing antimicrobial peptides or adjuvants with reduced experimental burden. The reviewed evidence indicates that AI substantially enhances decision making in scope of AMR challenges. Effectively continued impact will depend on data quality, model development and integration into public health and laboratory infrastructures.

Introduction

Antimicrobial resistance (AMR) is the ability of infectious pathogens to become resistant to one or more antibiotics, making infections increasingly difficult to treat and posing a serious global health threat. Over the past decade, AMR has increased significantly, primarily due to the overuse and misuse of antibiotics in healthcare, agriculture, and animal husbandry. AMR continues to spread worldwide due to limited or ineffective treatment options, poor infection control measures, and inadequate sanitation in some community and healthcare settings.^1,2

The clinical burden of AMR is substantial with he level of resistance varying among pathogens, with multidrug-resistant and pan-resistant organisms posing major concerns with notable examples like the cephalosporin-resistant Escherichia coli, methicillin-resistant Staphylococcus aureus (MRSA), and extended-spectrum β-lactamase (ESBL) producing Klebsiella pneumoniae.^3–5 These pathogens have developed into a global health crisis, leading to treatment failure, prolonged hospital stays, higher death rates, and a significant economic burden. ⁶ In 2019, bacterial antimicrobial resistance (AMR) was associated with about 4.95 million deaths globally, including 1.27 million deaths directly caused by resistant infections. The highest death rates were observed in western sub-Saharan Africa. Staphylococcus aureus, accounted for most AMR deaths, with methicillin-resistant S. aureus alone causing over 100,000 deaths worldwide. ⁷ Therefore, enhanced diagnostic and surveillance approaches are urgently needed to address this growing threat of AMR. Current detection systems emphasize rapid point-of-care (POC) diagnosis to reduce inappropriate use of broad-spectrum antibiotics, ⁸ their widespread implementation is limited due to their high costs and the need for specialized expertise. Addressing these barriers is essential for the United States National Strategy for Combating Antibiotic Resistance, which prioritizes improving access to effective diagnostics, promoting the development of new antibiotics, and optimizing the use of existing antimicrobials. ⁹ In laboratory settings, the common approaches for Antimicrobial Susceptibility Testing (AST) techniques vary from one setting to another, such as, oxford cup assay, broth dilution, agar dilution, disk diffusion, and gradient strip. Additional more advanced techniques include immunoassays and enzymatic assays, polymerase chain reaction (PCR) for genotypic approaches, which relies on genome sequencing and metagenomics.^10–12 Techniques that rely on built-in instrumental artificial intelligence and smart technologies such as matrix-assisted laser desorption or ionization, time-of-flight mass spectrometry (MALDI-TOF MS), microfluidic-based techniques, and automated systems such as BD Phoenix Microbiology System (BD Diagnostics, USA) and Sensititre ARIS 2X (Trek Diagnostic Systems, USA), offer real time large batches of sample testing and reduced human error. However, most of these techniques still require pure cultures and technical expertise. ¹³

Surveillance is a cornerstone in AMR management by monitoring resistance patterns and mitigating further resistance development across clinical, environmental, and agricultural domains.^2,14,15 Surveillance data are monitored nationally and internationally by organizations such as the Global Antimicrobial Resistance Surveillance System (GLASS) established by the World Health Organization (WHO). GLASS encompasses various components; including antimicrobial consumption monitoring (AMC), fungal pathogen surveillance (GLASS FUNGI), and emerging AMR reporting (GLASS EAR). ² GLASS collaborates with regional networks and national programs, such as the European Antimicrobial Resistance Surveillance Network (EARS-NET), a publicly funded AMR surveillance system that facilitates continuous monitoring and data sharing among policymakers.^16,17 Despite these frameworks, existing surveillance systems faces inherent limitations, such as geographic coverage, time of action and the integration of heterogeneous data sources. Nevertheless guidelines issued by comprehensive surveillance systems can significantly contribute to addressing AMR challenges. ¹⁷

Artificial Intelligence (AI) has transformed numerous aspects of healthcare delivery. AI applications support AMR diagnostics, personalized medicine, medical imaging, data analysis, and predictive modeling. Despite ethical considerations and accuracy challenges, the advantages of AI implementation are revolutionary.^18–20 Integrating AI into healthcare systems is projected to save approximately $150 billion of healthcare costs in the United States alone. AI applications can be categorized into four domains: augmented intelligence and assisted intelligence, which support clinical decision making, and automated intelligence and autonomous intelligence, which handle routine administrative and operational tasks with minimal human involvement.^21,22 The common used healthcare AI methodologies incorporate; machine learning (supervised, unsupervised, and reinforcement learning algorithms), natural language processing, deep learning, neural networks, and machine vision.^23–25 Indeed, AI can support the surveillance systems through processing of genomic data across diverse bacterial populations.^26,27 This review explores AI applications in AMR, across four thematic domains: 1) phenotypic and genotypic bacterial identification 2) antimicrobial susceptibility testing 3) AMR Surveillance 4) antibiotic development, providing a consolidated resource for researchers, clinicians and policymakers to work towards harnessing the power of AI in the fight against antimicrobial resistance.

Methods

This review employed a structured, comprehensive search of PubMed, Web of Science, and Google Scholar, with no language or year restrictions. The search was conducted on January 30, 2025, using the MeSH terms of “Artificial Intelligence” in combination with the “Antimicrobial Resistance”, or “Surveillance”, or “Antibiotics” or “diagnosis” or “stewardship”. We included related peer reviewed journal articles, reviews and conference proceedings reporting applications of AI in AMR diagnostics or surveillance. Citations retrieved from the search strategies were imported into EndNote (version 20.2.1) for screening and narrative synthesis for relevance which. Were mapped and summarized into four themes related to AI in AMR Diagnostic Reporting was guided by the Scale for the Assessment of Narrative Review Articles (SANRA). ²⁸

AI in diagnostics of AMR

Phenotypic and genotypic identification

Traditional bacterial classification based on phenotypic characteristics has been the predominant methodology which underpins all downstream clinical decisions in AMR management, including treatment selection and infection control. Although this approach is widely available, are prone to human error and time consuming methods while genotypic approaches offer better precision but have been constrained by cost and technical expertise. Several studies demonstrate that artificial intelligence (AI), including machine learning (ML) and deep learning (DL), enhances the accuracy, scalability, and reliability of infectious disease analysis compared with conventional manual approaches across both phenotypic and genotypic bacterial identification and resistance characterization. Machine learning (ML) offered bacterial species level classification by employing algorithmic models, such as the Support Vector Machine (SVM), Linear Regression, Decision Tree (DT), K-Nearest Neighbor (KNN), Neural Networks (NN), and Deep Learning (DL). The models are built on categorized data and training sets, followed by sequential stages of data processing, segmentation, feature extraction, and classification.^29,30

AI has been applied to the phenotypic characterization of tuberculosis (TB), where models learn complex patterns from data modalities such as; genomic sequences and medical images, thereby reducing dependence on human interpretation and improving the diagnostic workflows.^31–33 In the context of Mycobacterium tuberculosis (MTB) phenotyping, the genome-based ML approaches have shown strong capability in predicting drug resistance patterns using whole genome sequencing data. Ensemble and tree-based ML models also enabled resistance prediction while identifying rare or underreported mutations such as the loss of function variants. Such achievement highlight the value of large genomic datasets to improve robustness and generalizability.^31,32 Beyond genomics, the deep learning approaches that were applied to chest X-ray images have demonstrated effectiveness in detecting drug resistant tuberculosis, providing a complementary phenotypic assessment at the radiological level. Such imaging based AI systems can serve as screening and triage tools, particularly in resource limited settings. ³³

Artificial intelligence (AI) applications in microbiological imaging have further expanded to address challenges associated with manual interpretation and labor intensity. The applications of AI in microbial imaging included bacterial identification, colony detection and counting, Gram-stain interpretation, antimicrobial susceptibility testing, and clinical diagnostic support.^34–37 Several studies reported the reliance on convolutional neural networks (CNNs) to extract complex visual features from microbiological images.^38–42 The studies differ in diagnostic scope, microbiological context, dataset scale, and AI task formulation. These differences highlight the adaptability of deep learning architectures to diverse microbiological workflows.

For instance, Zieliński et al. ³⁸ presented one of the earlier comprehensive applications of deep learning for bacterial colony classification using microscopy images. Using CNN based feature extraction combined with classical classifiers (Support Vector Machine and Random Forest), the study demonstrated robust discrimination across 33 clinically relevant bacterial genera such as Escherichia coli, Staphylococcus aureus, Pseudomonas aeruginosa, and Enterococcus spp. Building on colony level analysis, Majchrowska et al. ³⁹ introduced the AGAR dataset, one of the largest publicly available resources for microbial colony detection and counting. Using Faster R-CNN and Cascade R-CNN architectures, the study focused on object detection rather than taxonomic classification. The dataset included colonies of Staphylococcus aureus, Bacillus subtilis, Pseudomonas aeruginosa, Escherichia coli, and Candida albicans grown on agar plates under diverse imaging conditions. The study demonstrated that deep neural networks can support laboratory automation and reducing dependence on manual colony enumeration. Consistently, Ferrari et al. ³⁷ addressed bacterial colony counting within the context of digital microbiology imaging and full laboratory automation systems. The study reported that deep learning markedly improved performance, particularly for complex colony aggregates when comparing CNN-based methods with manual classifiers. Although species identification was not the primary goal, the study reinforced the robustness of CNNs for quantitative microbiological tasks across diverse clinical samples.

Beyond colony level imaging Smith et al. ⁴⁰ applied a deep CNN to the automated interpretation of Gram stained blood culture smears. Using tens of thousands of images derived from routine clinical samples, the model classified Gram-positive cocci in clusters, Gram-positive cocci in chains or pairs, and Gram-negative rods, representing pathogens such as Staphylococcus spp., Streptococcus spp., and common Gram-negative bacilli. The CNN achieved high accuracy and strong generalization at the slide level. Additionally, AI microbiological diagnostics are further examined by Wang et al., ³⁵ who developed NugentNet, a specialized CNN for automated diagnosis of bacterial vaginosis (BV) from Gram-stained vaginal smears. The model quantified morphotypes corresponding to Lactobacillus spp., Gardnerella vaginalis, and Mobiluncus spp., enabling automated Nugent scoring. The CNN demonstrated superior accuracy, consistency, and reproducibility compared with trained technologists, supporting its clinical applicability in gynecological microbiology.

Pizzato et al. ⁴³ and Hsiao et al. ⁴² collectively illustrate the expanding scope of artificial intelligence in microbiology beyond conventional classification toward quantitative and biochemical pattern recognition. Pizzato et al. ⁴³ demonstrated that machine learning applied to MALDI-TOF MS lipid profiles enables reliable discrimination of Escherichia coli, Shigella flexneri, and Shigella sonnei, species that are difficult to distinguish using routine approaches. Complementarily, Hsiao et al. formulated bacterial colony enumeration as a supervised regression problem, introducing a patch based AI framework for accurate CFU quantification of E. coli from Petri dish images, achieving high accuracy. Together, these studies highlight a methodological convergence in which AI shifts microbiological analysis from categorical identification alone toward data driven quantitative inference, whether through lipid signatures or regression cytometric estimation.

Bacterial growth monitoring has also been addressed using artificial intelligence (AI) to support routine clinical bacteriology workflows. Jacot et al. ⁴⁴ developed and evaluated an algorithm for automated detection of bacterial growth and early growth monitoring within laboratory automation environment. The system was trained and validated using 3,844 culture plates derived from a broad range of clinical specimens and five common agar media, with plates imaged repeatedly at up to 15 incubation timepoints between 1 and 48 h. Notably, the system detected bacterial growth earlier than conventional manual reading. The AI model identified colony forming unit (CFU) candidates and analyzed temporal changes, which achieved sensitivity of 99.80 % and specificity of approximately 92 % at 48 h. Overall, this study highlights the value of AI growth monitoring as a scalable tool for enhancing efficiency in modern clinical bacteriology laboratories. Table 1 represents Performance comparison of example AI models for AMR detection by application area.

Table 1.

Performance comparison of example AI models for AMR detection by application area.

Application area	Top-performing AI models	AI/ML approach	Average performance	Time-to-result	Advantages	Limitations	References
Bacterial Identification	CNN-based Image Analysis, MALDI-TOF MS with AI, AI-assisted bacterial growth monitoring	Deep Learning, Machine Learning, Computer Vision	Accuracy: >90%	30 min - 2 hours	High accuracy, rapid results, minimal expertise required	May require specialized equipment, limited to known species	Zieliński et al. 38 ; Mortier et al. 45 ; Jacot et al. 44
Antimicrobial Susceptibility Testing	XGBoost models,	Machine Learning Ensemble, Integrated AI System, Deep Learning	AUC: 95%, Accuracy: >90%	2-4 hours	Significantly faster than traditional methods, automated interpretation	Limited clinical validation, may miss novel resistance mechanisms	Valavarasu et al. 46 ; Howard et al. 47 ; Noman SM 48
	Transformer-based Genomic Analysis, BioWeka (Java-based platform)
ARG Detection	DeepARG, ARG-BERT, Deep learning for metagenomic analysis	Deep Learning, Transformer-based	Precision: 0.95, Recall: 0.91	1-3 hours	High sensitivity for known ARGs, potential to detect novel ARGs	Requires high-quality sequence data, computational resources	Arango-Argoty et al. 49 ; Urel et al. 50
AMR Surveillance	AntiMicro.ai, Time series analysis for AMR, Network analysis for hospital transmission, ANNs model developed in SPSS Clementine for Methicillin-resistant S. aureus (MRSA)	Web-based ML platform, Deep Learning, Graph Neural Networks; Artificial Neural Networks (ANNs) model	Early warning accuracy: 87-92%	Real-time monitoring	Integrates multiple data sources, enables proactive interventions, personalized models	Complex implementation, requires data standardization	Mutisya & Kanguha 51 Kim et al., Chen et al.52,53; Myall et al.; Hsu CC 54
Point-of-Care Applications	AI-powered portable molecular diagnostics, Smartphone-based CNN, Smartphone-connected AI diagnostics	Machine Learning + Microfluidics, Deep Learning, Mobile Computing	Sensitivity: 88-92%, Specificity: 90-95%	<1 hour	Accessible in resource-limited settings, minimal training required	Lower accuracy than laboratory-based methods, limited pathogen range	Cavuto et al. 55 ; Banik et al. 56 ; Poirier et al. 57
Sepsis Prediction	BiLSTM with Attention, Personalized antimicrobial susceptibility models	Deep Learning, Machine Learning	AUROC: 0.91, Significant improvement in treatment outcomes	Real-time analysis	Early detection potential, integrates with EHR systems, personalized treatment	Requires comprehensive patient data, complex implementation	Arnold et al. 19 ; Howard et al. 58
Novel ARG Discovery	ARGminer, AMRFinder+, Deep domain adaptation	Machine Learning, Transfer Learning	Accuracy: 93%, Enhanced sensitivity for multiple markers	Variable	Potential to identify previously unknown resistance mechanisms	Limited validation of novel discoveries, false positive risk	Olatunji et al. 59 ; Mao et al. 60
Environmental Monitoring	WasteWater-ML, Climate-based AMR prediction, Wearable-based surveillance	Machine Learning, IoT + Machine Learning	Detection limit: 10 copies/mL, Enhanced early warning capabilities	1-2 days	Non-invasive surveillance, community-level monitoring, integration with climate data	Environmental variables may affect accuracy, standardization challenges	Li et al. 61 ; Nguyen et al. 34 ; Radin et al. 62
Clinical Decision Support	Personalized antimicrobial susceptibility models	Integrated AI System, Machine Learning	Significant reduction in inappropriate antibiotic use	Real-time to hours	Integration with clinical workflows, personalized treatment recommendations	Requires robust implementation and clinician acceptance	Howard et al.47,58; Bolton et al. 63
Automated Microbiology	AI-assisted bacterial growth monitoring, Deep learning for chromogenic media analysis	Computer Vision	Performance comparable to expert microbiologists, >95% accuracy	Hours vs. days for manual methods	Reduced workload, standardized interpretation, faster results	Initial implementation costs, requires validation against standard methods	Jacot et al. 44 ; Burns et al. 64 ; Faron et al. 65

Collectively, the research in this domain indicates that AI has brought a shift of microbiological identification from labor intensive, time consuming assays to data driven, quantitative inference. With the convergence of genomic, imaging and biochemical AI models can enable bacterial characterization. Although, most validated models are based on known species and curated datasets further work using diverse clinical samples with novels specimens can help in establishing a toolkit for detection.

AI applications in antimicrobial susceptibility testing

Antimicrobial susceptibility testing (AST) is the critical interface between diagnosis and treatment decisions. Accurate and rapid AST results can reduce the use of broad-spectrum agents, supports antibiotic prescription, and antimicrobial stewardship programs. Traditional techniques for AST are disc diffusion, broth microdilution which are time consuming techniques also requiring technical expertise. AI has been used extensively in AST to reduce manual workload, automate interpretation, and improve the reproducibility of readouts. For instance, Nguyen et al. ³⁴ developed an AI-powered disc diffusion reader that automates interpretation of Kirby-Bauer tests. Using Faster R-CNN for object detection and MobileNetV2 for antibiotic disc classification, the system analyzed AST plates containing Escherichia coli, Staphylococcus aureus, Pseudomonas aeruginosa, and Klebsiella pneumoniae. The model achieved high agreement with laboratory technicians while reducing inter observer variability. Gerada et al. ⁶⁶ developed AIgarMIC, a machine learning of agar dilution in which square plate photographs are split into sub images per inoculation spot and classified by a CNN in two stages (growth present vs absent, then inhibited vs good growth) to derive MIC endpoints and compare them against manual reading. This approach is valuable for large susceptibility data generation because it can operate using a consumer grade camera and standard computing hardware. Another initiative that focused on disc diffusion zone interpretation was guided by Bollapragada et al., ⁶⁷ which presented an AST “smart system” that uses CNN based semantic segmentation to identify inhibition zones on agar plate images, reporting strong agreement with traditional and manual readings and enabling fully automated inhibition zone measurement.

In contrary, AI encountered some limitations, Gullu et al. ³³ has developed an algorithm for inhibition zone detection from images with explicit practical challenges such as low contrast images and variable bacterial pigmentation under illumination. Similarly, Howard et al. ⁴⁷ reinforced the methodological distinction between disc methods (zone diameters) and MIC methods (minimum concentration preventing growth), underscoring different AI problem formulations. These advancements of AI applications in AST offer promising solutions to the traditional methodologies, potentially enabling more rapid and accurate determination of antimicrobial susceptibility, which is crucial for effective antibiotic stewardship and combating antimicrobial resistance. Across these studies, it has shown reduced assay turnaround time, improved interpretation and consistency of zone diameter and MIC. The ability to function with standard imaging hardware helps deployment of devices in low income countries. Hence, the outstanding challenge will be the regulatory validation of AI driven AST for adoption in accredited diagnostic laboratories.

AI applications in AMR surveillance

AMR surveillance is the systematic collection, analysis and reporting of resistance data across various settings to help guide public health policy and antibiotic stewardship. Traditional approaches faces constraints in reporting delays, fragmented data sources whereas AI plays a transformative role by enabling carrier screening, genomic resistance profiling, population level trend analysis and integration of heterogeneous data streams under One Health framework. In surveillance workflows, AI is frequently used to identify carriers or reservoirs through screening large volumes of cultures, or detect resistance from molecular data. A representative early clinical example is Hsu et al., ⁵⁴ who developed and validated an artificial neural network (ANN) to predict the likelihood of MRSA colonization using dozens of risk factors across two hospitals, reporting high accuracy with sensitivity and specificity balances that could support targeted screening decisions when universal surveillance is resource intensive. As per automated culture screening, Faron et al. ⁶⁵ evaluated Copan WASPLab automated digital analysis for vancomycin resistant Enterococcus (VRE), where the system captures images at time 0 after incubation, then scores plates by analyzing growth and colony color. The study illustrates a surveillance specific reality as automation may increase sensitivity which can also support infection control approaches.

Genomic and metagenomic surveillance provides a complementary pathway where AI helps interpret the resistance potential of sequenced data. Arango-Argoty et al. ⁴⁹ introduced DeepARG, proposing deep learning models that use a dissimilarity matrix across ARG categories to reduce false negatives associated with “best hit” database approaches, and reporting highly precise resistance categories. To support surveillance planning, more recent predictive modeling extends beyond single pathogens, but to include broad susceptibility prediction of large bacterial and fungal groups. Mutisya & Kanguha ⁵¹ reported decision tree based ML (XGBoost) achieving high AUC for antibacterial and antifungal susceptibility prediction and emphasized its feasibility through interactive web tools for clinician support.

The Big data driven AI analytics have been shown to strengthen population level surveillance by integrating heterogeneous datasets from national and international surveillance programs. Manik et al. ⁶⁸ demonstrated that machine learning models applied to large repositories such as WHO GLASS and national health records can uncover regional resistance trends and predict outbreaks. At the national policy level, AI surveillance has been integrated to planning. Odekunle et al. ⁶⁹ applied machine learning algorithms to multiyear surveillance data from Nigeria, combining predictive analytics with modelling in order to evaluate the impact of stewardship programs and public education on AMR. This work illustrates how AI surveillance can move beyond monitoring to actively inform policy design and resource allocation, particularly in low and middle income settings. In addition, the One Health surveillance represents a natural extension of AI in AMR by integrating data across human, animal, and environmental reservoirs, thereby building on surveillance approaches that leverage clinical records and genomic information. AI driven One Health systems apply ensemble machine learning models to analyze heterogeneous data from different resources which enable early detection of resistance emergence and geographically distributed resistance signals. ¹⁴ Biosensors AI integration within One Health surveillance architectures has been used as well. Lawal et al. ⁷⁰ reviewed emerging AI biosensor that combine real time signal acquisition to detect resistance determinants across clinical, veterinary and environmental settings. These systems enable continuous data streaming and remote reporting. Therefore, these advances position AI as a key enabler in the field of policy relevant AMR surveillance at the global scale.

AI applications in antibiotic development

Antibiotic development has faced growing difficulties over the past decades, driven by high development costs and complex regulatory requirements especially for wet laboratory experiments. Artificial intelligence has emerged as a powerful tool to support antibiotic development efforts. AI driven approaches enable the rapid analysis of large chemical libraries and biological datasets, which help in identifying promising candidates more efficiently than traditional methods. Machine learning techniques are increasingly used to screen compound libraries and predict antibacterial activity and safety along with reducing experimental workload and development timelines. Through these contributions, AI supports not only the discovery of new antibiotics but also the use and extension of existing antimicrobial therapies.^4,63,71–74

In 2019, researchers utilized Deep Learning to screen a library of 107 million compounds, identifying a novel broad spectrum antibiotic effective against multidrug resistant pathogens. ²⁵ Studies designed novel antimicrobial peptides with specific activity against Gram-negative bacteria, achieving a 20-fold increase in discovery efficiency compared to traditional methods.^74,75 Generative Networks was used to design β-lactamase inhibitors demonstrating activity against carbapenemase-producing organisms. ⁷⁶ AI has also been applied to optimize the pharmacokinetic and pharmacodynamic properties of antibiotic candidates. Researchers employed Bayesian optimization to enhance the bioavailability of a promising antibiotic candidate, resulting in a 3-fold improvement in serum half-life. ⁷⁷ Studies utilized Quantitative Structure Activity Relationship models to predict and mitigate potential toxicity in a class of aminoglycosides, successfully reducing ototoxicity while maintaining antimicrobial efficacy.^78,79

Deep Learning model was used to optimize the tissue penetration of fluoroquinolone derivative, achieving improved concentrations in pulmonary tissue compared to existing agents.^71,80 Moreover, for combination therapy development, Random Forest algorithms have been employed to predict synergistic antibiotic combinations against multidrug resistant Acinetobacter baumannii with enhanced efficacy. ⁸¹ Other studies utilized Neural Networks to identify antibiotic adjuvant combinations that could restore susceptibility in carbapenem resistant Enterobacteriaceae, with promising activity in animal models.^72,73,82 An important milestone in AI antibiotic discovery was reported by Stokes et al., who trained deep learning models to predict antibacterial activity and applied them to screen large chemical libraries, resulting in the identification of halicin and other compounds with experimentally validated broad spectrum activity. ²⁵ Other studies emphasized on strategies to enhance or preserve the effectiveness of existing antibiotics. Zheng et al. ⁷² applied computational screening approaches to identify antibiotic adjuvants that restore activity against resistant pathogens rather than reinventing new antimicrobials. This work supported faster clinical translation by leveraging compounds that may already have known safety profiles.

A more targeted medicinal chemistry approach was presented by Aarjane et al., ⁷¹ who combined computational modeling with experimental synthesis to design new quinolone derivatives with improved antibacterial activity. Unlike broad library screening methods, this study focused on rational optimization within a defined chemical class. Beyond single agent optimization, combination based therapies represent another important direction in antibiotic development. Cantrell et al. ⁷³ examined machine learning to design antimicrobial combinations, addressing the complexity that makes experimental testing of all possible drug pairs impractical. The study emphasizes predictive modeling of synergy and dependent efficacy, positioning AI as a key enabler for designing combination regimens that can suppress resistance emergence while maximizing therapeutic benefit.

An additional extension of AI antibiotic development is the design of non-traditional antimicrobial agents. Fjell et al. ⁷⁴ focused on antimicrobial peptides, the study illustrated computational and data driven methods to learn relationships between peptides. Although peptide therapeutics differ from small molecule antibiotics in pharmacology and delivery, this study shares the same core objective as other AI approaches by navigating candidates that would be difficult to discover through trial alone. These AI approaches to antibiotic development offer potential solutions to the innovation gap in antimicrobial discovery, accelerating the identification of agents to address the growing challenge of antimicrobial resistance. ⁶³ Collectively, AI has been applied across the discovery to optimization stages, offering meaningful potential to reduce time and cost of bringing novel antimicrobials to clinical evaluation with the validation of AI discovered potential drugs in clinical trials remaining as the outstanding challenge. Figure 1 summarizes the transformative role of AI across the four thematic domains reviewed.

Figure 1.

Transformative role of AI in AMR diagnostics.

Challenges and social implications of AI applications on AMR

Generally, the application of artificial intelligence in antimicrobial resistance offers several opportunities and limitations for responsible deployment. The marked opportunities include strengthen diagnostics, surveillance, stewardship, and antibiotic development, yet several limitations include applicability, complexity, data privacy, generalizability issues, and technical challenges.^83,84 More specifically, one of the most significant opportunities lies in AI is the ability to analyze large and complex datasets that exceed human cognitive capacity. AI can uncover resistance patterns, predict emerging threats, and support evidence based decision making across clinical and public health settings.^29,85,86 AI also presents a major opportunity to expedite the labor intensive and time consuming tasks in laboratory routine testing. Automated interpretation of phenotypic tests, imaged diagnostics, and susceptibility assays improves consistency and scalability while reducing reliance on specialized expertise. ⁸⁵ Similarly, in surveillance, AI support screening and predictive modeling, allowing resistance monitoring and early warning.^29,86

Despite the substantial progress of AI in AMR, there are also several challenges which hinder its real-world impact. The major challenges are data heterogeneity and lack of standardization among various healthcare systems and laboratories. However, AI models are trained on curated datasets which do not represent the variability of data encountered in real-world clinical settings. Therefore, the different diagnostic systems, reporting standards and AST protocols limit model generalizability especially among low and middle income countries. ⁸⁷

The AI model also require continuous updates and calibration as resistance patterns can evolve. Although there are various AI tools developed with high performance in pilot or retrospective studies, with limited clinical validation and regulatory oversight preventing it to be clinically adopted. Even though regulatory frameworks are constantly evolving with concerns of algorithm transparency, accountability and explainability when AI can play a role in clinical decision making. ⁸⁵ The black box nature of models can further reduce clinical trust and hinder integration into systems. Therefore, technical and regulatory constraints affect the adoption of AI models.

Ethical and behavioral factors also represent a critical limitation. These AI driven systems cannot be used autonomously as they can prioritize population level optimization over individual patient welfare. Rather, AI should be augmentative, inform and support in decision making without overriding patient autonomy. Therefore, AI supported systems can enhance diagnostic confidence, identify resistance risks and provide evidence based recommendations without interfering with clinician autonomy and patient choice. ⁸⁸ Hence, AI can help reduce the development of AMR with ethical principles in check with transparent outputs, uncertainty estimates and clinician oversight mechanisms.

Antibiotic overuse can also be affected by decision making involving patients, clinicians and health systems. Along with the idea of optimal antibiotic prescription must be viewed as context dependent due to the ethical issues and incentive imbalances rather than an AI criteria. Patients prefer antibiotic use for rapid recovery even while clinical benefit is moderate meanwhile physicians despite awareness of AMR, may prescribe antibiotics due to pressure, institutional incentives, and anticipated patient dissatisfaction. ⁸⁹

AMR can be viewed as a social conundrum where decisions that are rational at an individual level can conflict with outcomes that are best for society. ⁹⁰ Even though antibiotic use is restrained by others, individuals can still benefit and in turn the harmful consequences are delayed, shared and often borne by the population at large, across clinical practice, agriculture and international settings. ⁹¹ Hence, coordinated AMR mitigation strategies can be challenging to sustain, when responsible conduct is insufficiently encouraged and non-compliance not resulting in immediate penalties. While AI can help reduce discovery costs and accelerate antibiotic development, incentive misalignments and market constraints contribute to persistent decline in research and development. ⁹² But with AI and economic policies, public private partnerships can accelerate early stage development of new age antibiotics which are often reserved as last line therapy. These challenges and limitations highlight the fact that AI cannot address AMR crisis alone but with alignment of ethical governance, regulatory supervision and incentive systems that support responsible antibiotic use.

Discussion

This narrative review has delved into four thematic domains where in AI has effectively shown meaningful contributions to AMR diagnostics and management. The convergent picture that emerges from phenotypic and genotypic identification, AST, surveillance and antibiotic development is that AI can improve various aspects of assays such as time, accuracy and scalability. AI helps to overcome the limitations of human expertise, data processing capacity and laboratory throughput. The findings of this review are in line with similar systematic and scoping reviews such as Abu El Ruz et al.^83,84 have listed AI application for bacterial infection control, along with identifying diagnostic automation and limitations in clinical validation. Howard et al. ⁸⁵ and Dien Bard et al. ⁸⁶ have analysed and concluded that although diagnostic AI for clinical microbiology has achieved technical precision, the transition from research demonstration to clinical deployment remains a challenge. This review also corroborates these facts and provides thematic evidence supporting them.

A recurring observation across all the four themes has shown the AI models have been evaluated in controlled settings with a performance gap between those deployed in multi center environments. Most of the studies have relied on a single institutional datasets or publicly available datasets. Therefore, the generalizability of these models to the diverse clinical population can have profound implications especially in AST and surveillance with substantial variations across health systems. The One Health principle being integrated into AI driven AMR surveillance is a promising approach to identify resistance emergence before they manifest in human clinical populations. The proactive approach aligns with AMR action plan and showcases predictive resistance monitoring among clinical, veterinary and environmental data streams with unified machine learning frameworks. However, the One Health AI surveillance will also require large investments to develop the frameworks and infrastructure. While in antibiotic development the identification of halicin by Stokes et al. ²⁵ appears to be the proof of concept for AI generated antibiotic candidate. This compound has not yet been clinically approved which appears to be a challenge even though AI accelerates the initial stages of identification. These AI generated candidates will also need to be compared to the conventionally identified compounds for efficacy.

There are several limitations to this review with rapidly evolving AI based literature and studies being published and not being incorporated. The current review has also not been stratified based on geographic or resource settings as it can show disparities in AI applicability between developing and developed countries. Despite these limitations, this review covers the four themes comprehensively for AI applications across AMR diagnostics. The various studies are indicative of the fact that AI is no more a speculative technology but an integral part of modern diagnostic and surveillance. Furthermore, prospective models require further validation, data standardization and regulatory frameworks for deployment.

Conclusion

The body of evidence reviewed demonstrates that artificial intelligence has become an integral component of modern strategies to address antimicrobial resistance, with meaningful impact across diagnostics, susceptibility testing, surveillance, and antibiotic development. In bacterial identification, AI enables reliable phenotypic and genotypic characterization by learning complex patterns from genomic data, microscopy, culture images, and biochemical profiles, reducing dependence on subjective interpretation and specialized expertise while improving diagnostic consistency and scalability.

In antimicrobial susceptibility testing, AI driven approaches have standardized the interpretation of disc diffusion and MIC based assays through automated image analysis and machine learning classification. These systems reduce inter observer variability and accelerate result generation, strengthening the reliability of susceptibility data that inform clinical treatment decisions and antimicrobial stewardship. At the surveillance level, AI has shifted resistance monitoring from retrospective reporting toward predictive and integrative frameworks. AI supports early detection of resistance trends, targeted screening, and policy relevant forecasting by analyzing large scale clinical, genomic, and environmental datasets, including One Health surveillance models that capture transmission across human, animal, and environmental reservoirs.

In antibiotic development, AI addresses the long standing innovation gap by enabling efficient exploration of chemical and biological space. Machine learning models support the discovery of novel antibacterial agents, the rational optimization of existing compounds, the design of antimicrobial peptides, and the identification of effective drug combinations and adjuvants. These approaches reduce experimental burden and accelerate progression from candidate identification to therapeutic evaluation. The real world translation of these stated capabilities into routine clinical practice will require further improvements in data quality, multi-center validation of AI model, regulatory procedures and sustained investments for deployment in digital health infrastructure in low income countries where AMR is a significant burden. AI is not a solution to AMR but with time and appropriate governance, validation and integration, it is one of the most powerful tools to this critical public health challenge.