Artificial Intelligence in Rapid Diagnostics for Antimicrobial Resistance: From Bench to Bedside

ABSTRACT

Background

Antimicrobial resistance (AMR) is one of the most urgent global public health threats, especially in low‐ and middle‐income countries. In 2019, bacterial AMR was estimated to have directly caused approximately 1.27 million deaths and to have been associated with 4.95 million deaths worldwide. Rapid identification of resistant infections is important, yet established diagnostics are very slow. In recent years, artificial intelligence (AI) has revolutionized microbiology and infectious disease diagnostics. This review explores the space of AI‐enabled diagnostics for AMR detection and characterization.

Methods

This literature review was conducted from January to September 2026 using PubMed/MEDLINE, Google Scholar, and other electronic databases with keywords including “AI,” “AMR,” “deep learning,” “genomic diagnostics,” and “point‐of‐care (POC) testing.” Following screening and full‐text assessment, 67 Peer‐reviewed articles and systematic reviews were included to support the synthesis presented in the manuscript, while non‐peer‐reviewed articles and studies lacking direct AI relevance were excluded. Titles and abstracts were screened before full‐text review.

Results

AI‐enhanced diagnostics were reported to shorten turnaround time for AMR detection with promising but heterogeneous predictive performance; for example, pooled estimates across antimicrobial stewardship applications indicate an area under the curve of approximately 0.72. They also support improving antimicrobial stewardship and enable early therapeutic optimization in laboratory and POC settings, enabling real‐time clinical decision‐making. Despite these achievements, evidence for direct improvements in patient outcomes and routine bedside effectiveness remains limited.

Conclusion

AI‐driven diagnostic tools represent a promising strategy to bridge laboratory innovation and bedside application in combating AMR. The tools have the capacity to improve patient outcomes and enhance antimicrobial stewardship as they can accurately detect resistance patterns and support real‐time clinical decisions. However, long‐term sustainability, long‐term success, and global impact can be achieved by continued validation, regulatory alignment, and equitable implementation across diverse healthcare settings.

Abbreviations

AI

artificial Intelligence

AMR

antimicrobial resistance

AST

antimicrobial susceptibility testing

CNN

convolutional neural network

DL

deep learning

LMICs

low‐ and middle‐income countries

MALDI‐TOF

matrix‐assisted laser desorption/ionization time‐of‐flight

MICs

minimal inhibitory concentrations

ML

machine learning

POC

point‐of‐care

SNPs

single nucleotide polymorphisms

WGS

whole‐genome sequencing

WHO

World Health Organization

1. Introduction

Antimicrobial resistance (AMR) is one of the greatest global public health threats. In 2019, bacterial AMR was estimated to have directly caused approximately 1.27 million deaths and to have been associated with 4.95 million deaths worldwide [1]. Current trends are alarming and need urgent intervention, as AMR is projected to result in approximately 10 million deaths annually by 2050, alongside cumulative economic losses of up to 100 trillion US dollars [2]. These figures reflect how resistance undermines treatment of common infections and jeopardizes advanced medical procedures, especially in low‐ and middle‐income countries (LMICs) where surveillance and healthcare resources are limited [2, 3]. Rapid identification of resistant infections is important, yet conventional diagnostics are very slow. Conventional microbiology processes are based on culture and phenotypic analysis: species identification by culturing a pathogen can require a few days to weeks, and an additional 1–2 days are spent on susceptibility testing [4]. Therefore, there is a dire need for much faster and more accurate diagnostics.

New rapid tests are being developed and could revolutionize infectious disease care by providing timely results and conserving effective antibiotics [5, 6]. Faster identification of pathogens and their resistance profiles is necessary to guide appropriate therapy, enabling appropriate antimicrobial selection and de‐escalation, but the extent to which faster diagnostic information translates into improved patient outcomes depends on clinical implementation and validation [7]. In recent years, there has been development of artificial intelligence (AI) to revolutionize as a tool in microbiology and infectious disease diagnostics. AI‐based approaches have the potential to identify patterns or signs of resistance using large and complex data sets, including machine learning (ML) and deep learning (DL), which can identify species, predict resistance phenotypes, and even anticipate outbreaks by finding subtle correlations in data that can't be determined by human analysis [8]. A study by Pinto et al. (2025) demonstrated AI applications across infectious disease diagnosis, prognosis, therapeutic decision support, and antimicrobial stewardship, indicating that the role of AI extends beyond laboratory classification to broader clinical and public health applications [9].

In older adults, AI was used for the diagnosis of infectious disease, prognosis, and therapeutic decision support. The study highlights both the potential of AI and the importance of validation specific to the population, particularly in populations characterized by multimorbidity, polypharmacy, and potentially atypical clinical presentations [9]. For instance, ML models trained on Matric‐assisted lased desorption/ionization time‐of‐flight mass spectrometry (MALDI‐TOF) spectra or genomic features have predicted resistance phenotypes earlier than conventional phenotypic AST. However, model inference time should not automatically be equated with total sample‐to‐result time, as it depends on specimen processing, culture or sequencing, data acquisition, and clinical workflow [10, 11].

Such AI‐driven tools can help close the gap between rich laboratory data and clinical decision‐making [10]. This review explores AI‐enabled AMR diagnostics across the bench to bedside continuum, including how AI (especially ML/DL) is applied in microbiology workflows (e.g., MALDI‐TOF, sequencing, automated imaging) and in point‐of‐care (POC) platforms, and how these tools can shorten time‐to‐result and improve antimicrobial stewardship. It also presented challenges, limitations, and ethical considerations, as well as future directions and research gaps. As noted in recent literature, ML will transform AMR detection and treatment optimization [2]; here, we focus on the “bench‐to‐bedside” journey of AI diagnostics for AMR.

2. Methods

This literature review was conducted between January and September 2026 using PubMed/MEDLINE and Google Scholar. Search terms included AI, AMR, deep learning, genomic diagnostics, molecular diagnostics, MALDI‐TOF, and POC testing. Boolean operators (AND, OR) were used to combine search terms. The searches retrieved peer‐reviewed articles and systematic reviews. Following screening and full‐text assessment, 67 Peer‐reviewed articles and systematic reviews were included to support the synthesis presented in the manuscript, while non‐peer‐reviewed articles and studies lacking direct AI relevance were excluded. Exclusion criteria included non‐peer‐reviewed articles and studies lacking direct AI relevance. Titles and abstracts were screened before full‐text review.

As a narrative review, no formal risk of bias or standardized quality appraisal tool was applied. Literature was assessed based on relevance, methodological characteristics, reported diagnostic or predictive performance, and clinical applicability, and findings were synthesized narratively. Attention was given to differences between retrospective, experimental, non‐validated models and evidence from prospective or clinically validated applications. Based on the level of evidence and clinical readiness of the underlying studies, claims regarding diagnostic accuracy, turnaround time, antimicrobial stewardship, and clinical decision‐making were interpreted.

3. Results

3.1. Overview of Artificial Intelligence Techniques Used in AMR Diagnostics

3.1.1. Evidence Level and Clinical Readiness

AI in healthcare generally refers to computer systems performing tasks that typically require human intelligence (see Figure 1) [12]. The evidence supporting AI‐enabled AMR diagnostics is different in terms of methodological maturity and clinical readiness. Applications discussed in this review are first clinically evaluated applications, in which performance has been assessed against established reference methods or clinical datasets. Secondly, early‐stage applications or experimental applications were evaluated using retrospective, laboratory‐based, single‐Center or preclinical datasets. Lastly, emerging or future applications for which clinical utility remains to be established. Recent evidence syntheses, including a systematic review and meta‐analysis of 80 studies, have reported promising predictive and diagnostic performance of AI applications in antimicrobial stewardship, but also substantial heterogeneity and limited external validation [8, 11, 13]. Reports show reductions in turnaround time or improvements in predictive accuracy and are interpreted based on the evidence setting. Computational inference time is distinguished from demonstrated improvements in antimicrobial, patient outcomes, or routine bedside decision‐making [8, ⁹].

Figure 1.

AI workflow for AMR diagnostics: data input → model training → prediction → clinical decision.

3.2. Data Inputs

AI models are capable of consuming various data types that are applicable to AMR (see Figure 1). To illustrate, MALDI‐TOF mass spectrometry produces high‐dimensional spectral fingerprints; the ML/DL models can be trained on the spectra to predict resistance [10]. Whole‐genome sequencing (WGS) data give the characteristics of the presence/absence of genes or single‐nucleotide variations; ML techniques apply such genomic characteristics to determine the susceptibility profiles [14]. CNNs can process microscopy or digital imaging data (e.g., time‐lapse images of bacterial growth on agar or in microfluidics) to recognize species and even identify morphological changes that are linked to resistance [15]. AI models can also be fed with other inputs, such as Raman or infrared spectra of bacterial samples [16], metabolomic profiles, or even patient clinical data (demographics, medication exposure, comorbidities, lab results, antibiotic history) [17]. The nature of the input data guides the modeling approach: image and spectral data often use DL, whereas tabular genomic or clinical data may use classical ML [18].

3.3. AI Applications in Microbiology Laboratories

3.3.1. AI‐Enhanced MALDI‐TOF for AMR Detection

Matrix‐assisted laser desorption/ionization time‐of‐flight mass spectrometry (MALDI‐TOF MS) has become an essential tool in clinical microbiology laboratories due to its rapid, accurate, and cost‐effective bacterial species identification [19, 20]. Compared to conventional techniques, including antimicrobial susceptibility testing (AST) that depends on slow culture‐based methods, MALDI‐TOF can provide results and identify microorganisms in minutes. Recent studies have shown that integrating AI with MALDI‐TOF spectral data can enable rapid prediction of AMR, significantly shortening diagnostic timelines [21, 22].

Recent studies show that AI/ML applied to MALDI‐TOF spectra helps in predicting resistance earlier than conventional phenotypic AST in selected study settings. Nevertheless, most evidence is mainly based on clinical or laboratory datasets rather than prospective demonstrations of improved patient outcomes; therefore, interpretation of rapid model‐based prediction should not be automatic as a universally shorter sample‐to‐result diagnostic pathway [8, 21, 22].

As illustrated in Figure 2, the typical AI‐enhanced MALDI‐TOF workflow involves acquisition of protein spectra during routine species identification, spectral preprocessing, and subsequent analysis using ML or DL models to generate resistance predictions. These models detect subtle multidimensional patterns in spectral data that correlate with resistant phenotypes, enabling probabilistic predictions of AMR before completion of conventional AST [21, 23]. This approach is feasible across a range of clinically relevant pathogens. Weis et al. analyzed a large collection of clinical MALDI‐TOF spectra linked to corresponding AST results, showing that ML models could predict resistance phenotypes for Staphylococcus aureus, Escherichia coli, and Klebsiella pneumoniae with good discriminative performance (AUROC ~ 0.74–0.80) in a large clinical dataset [21].

Figure 2.

AI‐enhanced MALDI‐TOF mass spectrometry workflow for rapid antimicrobial resistance prediction (adapted from Weis et al., 2022; Weis et al., 2023; Nguyen et al., 2023; Smith et al., 2024; Tamma & Simner, 2020).

A validation study further showed that models such as random forests and support vector machines maintained predictive accuracy (AUROC ≥ 0.8) when trained and periodically retrained on local clinical data. However, their performance declined when applied to temporally distant datasets without adjustment [24]. A similar model trained on MALDI TOF spectra of Pseudomonas aeruginosa demonstrated strong performance (AUROC ~ 0.87) for predicting resistance to β‐lactam/β‐lactamase inhibitor combinations [25]. Furthermore, a study conducted on K. pneumoniae further identified specific MALDI‐TOF spectral features associated with multidrug resistance phenotypes, supporting model interpretability [23]. In addition, a systematic review confirms that ML methods applied to MALDI‐TOF spectra can enhance rapid resistance prediction across diverse bacterial species [22].

As summarized in Table 1, these studies used different algorithms, including random forests, support vector machines, and convolutional neural networks, and report substantial reductions in time to resistance prediction compared with conventional AST. AI‐enhanced MALDI‐TOF represents a scalable, cost‐effective approach for early AMR detection within routine laboratory diagnostics. However, site‐specific validation and model retraining remain essential. Evidence concerning cost‐effectiveness, scalability across laboratories, as well as prospective clinical impact remains comparatively limited [8].

Table 1.

Key studies using AI with MALDI‐TOF for AMR detection.

Pathogen	Algorithm	Accuracy/AUROC	Evidence/clinical readiness	Reference
Staphylococcus aureus	Random Forest, SVM	AUROC 0.74‐0.80	Clinical dataset; early‐stage predictive application	[21]
Escherichia coli	ML models	AUROC 0.78	Clinical dataset; early‐stage predictive application	[21]
Klebsiella pneumonia	Informative spectral Peaks + AI	AUROC 0.81	Study‐specific validation, early stage	[23]
Pseudomonas aeruginosa	Random Forest/ML	AUROC 0.87	Clinical MALDI‐TOF dataset; early stage	[25]
Various Bacterial Species	Machine Learning approaches	Variable	Systematic review evidence, heterogeneous validation	[22]

3.4. AI and Genomic‐Based AMR Diagnostics

WGS is an important part of AMR prediction, which enables comprehensive analysis of bacterial genomes to identify resistance determinants and infer phenotypes directly from DNA sequence data. AI‐driven genomic approaches don't rely on prior knowledge of known resistance genes for targeted gene detection. They leverage ML models to detect complex genomic patterns, such as combinations of single‐nucleotide polymorphisms (SNPs), gene presence/absence, and k‐mer features, that are associated with resistance phenotypes [14]. These models can incorporate large datasets to predict antimicrobial susceptibility with high accuracy and scalability [26].

Studies that employed WGS data to predict resistance have revealed that models such as random forests, support vector machines, and convolutional neural networks can successfully infer resistance phenotypes for multiple antibiotics across diverse pathogens [27]. For example, ML models trained on large datasets of Salmonella genomes attained high accuracy in predicting minimal inhibitory concentrations (MICs) without relying on prior gene annotations, showing the potential of data‐driven genomic prediction to augment traditional AMR diagnostics [27]. In Acinetobacter baumannii, models built from k‐mer features extracted from WGS data achieved prediction accuracies exceeding 90%, demonstrating rapid MIC inference that could be completed within minutes once the genomic data are available [26].

Integrative approaches that combine gene, SNP, and pan‐genome features have further improved prediction performance and helped identify genomic markers that contribute strongly to resistance, as shown in studies of Staphylococcus aureus using WGS data across thousands of isolates. Additionally, as shown in Table 2, ML frameworks have been used to quantify the contribution of AMR‐related SNPs in Mycobacterium tuberculosis, enhancing the interpretability of genomic predictions and enabling more transparent models for clinical use [28]. Culture‐independent metagenomics expands this paradigm by enabling resistance prediction directly from clinical or environmental samples without prior cultivation, using mNGS to detect AMR gene signatures across microbial communities [29]. As summarized in Table 2, AI‐driven genomic tools vary in data source, prediction targets, and strengths, such as high‐dimensional pattern recognition, as well as limitations, including computational requirements and the need for extensive training datasets.

Table 2.

AI‐driven genomic tools for AMR prediction.

Tool/Approach	Data source	Prediction target	Strengths	Limitations
WGS ML models (Salmonella)	Whole genome sequences	MIC/resistance phenotype prediction	Genome‐wide feature analysis	Needs large representative datasets
K‐mer ML (A. baumannii)	EGS k‐mers	MIC prediction	High‐dimensional pattern recognition	Computationally intensive, validation needed
SNP‐based ML(M. tuberculosis)	WGS SNP matrices	Resistance trait prediction	Identifies resistance‐specific markers	Requires curated datasets
Metagenomics AMR prediction	Metagenomics reads	AMR/resistance gene prediction	Culture‐independent	Lower specificity for phenotype

Together, these approaches in Table 2 demonstrate the potential of genomic‐based AI diagnostics to provide rapid, accurate insights into AMR and to support informed antimicrobial therapy. However, most available evidence remains retrospective or dataset‐specific; therefore, prospective external validation is required before routine bedside deployment [11].

3.5. AI in Automated Microscopy and Imaging

Automated microscopy combined with AI offers an emerging approach for rapid phenotypic antimicrobial susceptibility testing (AST) by directly observing bacterial responses to antibiotic exposure at the cellular level [30, 31]. Traditional phenotypic AST depends on visible population‐level growth changes, often requiring 18–24 h or more of culture. AI‐enhanced imaging accelerates this process by classifying subtle morphological and kinetic changes in single cells or microcolonies that happen soon after antibiotic treatment, aiding early discrimination between resistant and susceptible phenotypes [30].

Deep learning models, such as convolutional neural networks (CNNs), have been trained to identify antibiotic‐induced morphological differences and growth patterns in bacterial cells from time‐lapse microscopy images. A recent study showed that deep learning classifiers distinguished susceptible from resistant Escherichia coli cells treated with ciprofloxacin within as little as 30 min, showing the potential for single‐cell phenotypic AST with high concordance to clinical susceptibility definitions [30]. Moreover, earlier work using deep learning video microscopy demonstrated that automated analysis of bacterial motion and phenotypic features directly from urine samples could determine minimum inhibitory concentrations (MICs) within 30 min, offering rapid, culture‐independent susceptibility insights [32].

At subcellular resolution, structured illumination and other advanced imaging techniques combined with deep neural networks have delivered rapid AST results for multiple bacterial species, revealing cell heterogeneity that conventional methods mask and achieving agreement with reference AST results in minutes to hours [33]. Automated image analysis is also valuable for quantifying antibiotic‐induced morphological changes, such as filamentation, which can be an early marker of β‐lactam activity in E. coli, and can be quantified more objectively than manual interpretation [31].

Integrating AI‐driven image analysis with laboratory automation platforms enables high‐throughput detection and standardized interpretation, reducing observer variability and supporting large‐scale AST workflows. These systems are compatible with automated sample handling and microfluidic platforms, further increasing throughput and reducing time to result [34]. Overall, AI‐enhanced automated microscopy represents a significant route toward rapid, phenotypic AST and bridges the gap between lab automation and real‐time clinical microbiology, but prospective clinical validation is required to establish routine clinical utility and patient‐level benefit.

4. AI in Point‐of‐Care Testing for AMR

AI integrated into POC testing enables rapid testing of AMR directly at the bedside and near the patient [35]. Unlike laboratory‐based diagnostics, AI provides immediate resistance detection, enhancing precision in antimicrobial therapy. However, the presence of AI components does not by itself establish clinical validation or improved patient outcomes (see Table 3). Biosensors are analytical devices that detect AMR by using biorecognition elements and physicochemical transducers, which convert biological interactions into measurable signals. They include nanomaterials, which improve sensitivity and permit detection at low concentrations, and paper‐based and microfluidic designs, which support POC implementation. AI integration increases performance by reducing noise, recognizing complex patterns, and offering predictive insights into resistance trends. By uniting AI with biosensors, these systems could connect laboratory precision with real‐world applicability, contributing real‐time data‐driven surveillance that supports clinical decisions [36].

Table 3.

AI‐enabled POC AMR diagnostic tools.

AI/POC approach	Description	POC relevance	Evidence status	References
AI‐assisted smartphone biosensing	Image/signal analysis based on smartphone for rapid pathogen detection	Portable and potentially suitable for decentralized testing	Experimental/early stage	[37]
AI‐integrated biosensors	AI/ML may improve signal interpretation and resistance‐related pattern recognition	Potentially supports portable AMR detection and surveillance	Emerging	[36]
Raman spectroscopy assisted by machine learning	ML classification of Raman spectral signatures	Potential for rapid bacterial as well as resistance classification	Experimental/early stage	[38]
AI/ML diagnostic platforms	Computational models applied to microbiological and clinical data	Potential decision‐support role	Evidence varies and requires validation	[8]

Smartphones combined with microfluidic chips and advanced software are great tools for POC testing. They use ML and DL to process samples, capture images, and analyze data. Imaging methods like bright‐field and fluorescence, along with improved hardware, improve accuracy and usability. AI may help make diagnostics faster and more reliable [37]. Portable molecular platforms have revolutionized AMR detection by supporting rapid, on‐site diagnostics. CRISPR‐Cas sensors, aptamer‐based assays, and the Oxford MinION nanopore sequencer allow real‐time detection of resistance genes and pathogens outside the lab. These tools provide fast, accurate, and accessible alternatives to traditional laboratory methods, making molecular diagnostics more versatile [39]. CRISPR‐Cas assays, aptamer‐based platforms, biosensors and nanopore sequencing should, however, be distinguished from AI itself: these are molecular or analytical diagnostic platforms that may be combined with AI/ML for signal interpretation, classification, or downstream resistance prediction [39].

LMICs face major healthcare challenges due to limited infrastructure, high disease burdens, and economic constraints, especially in rural areas. POC diagnostics can improve access by enabling rapid testing and treatment outside central laboratories (see Figure 3). Their impact, however, depends on integration into local healthcare systems, supply chains, and workflows, as well as training of healthcare workers. Decentralized care has proven effective in overcoming these barriers and reducing loss to follow‐up [40].

Figure 3.

Bench‐to‐bedside pathway: lab‐based AI diagnostics to POC deployment.

4.1. Clinical Integration and Impact on Patient Care

AI integrated with POC diagnostics may improve AMR management by providing earlier predictive information and integrating microbiological and clinical data. (see Figure 4). However, evidence that AI can directly improve antimicrobial stewardship or patient outcomes at the bedside remains heterogeneous. It may support patient‐specific clinical decisions by integrating microbiological and support analysis of complex clinical and laboratory data to offer real‐time, patient‐specific therapeutic recommendations, helping clinicians select the most appropriate antimicrobial therapy faster than conventional methods. However, evidence of improved clinical outcomes depends on the model, population, implementation setting, and degree of validation [9, 41]. ML may improve AMR predictions and optimize antibiotic prescribing, enhancing clinical decision accuracy when guiding therapy choices at the bedside [42]. ML has demonstrated promising predictive performance in antimicrobial stewardship, but predictive performance should not be equated with demonstrated improvement in prescribing or patient outcomes [8, ⁴¹].

Figure 4.

How AI‐enabled diagnostics inform antimicrobial prescribing and stewardship.

One of the most crucial advantages of AI with POC testing is shortening the time to appropriate therapy. AI‐guided systems enable rapid diagnostic outputs and support decision‐making, enabling clinicians to make decisions for antimicrobial therapies before obtaining conventional culture results (see Figure 4). AI may provide quicker, more accurate therapeutic recommendations and improve treatment precision [43]. AI‐supported systems may provide earlier diagnostic or therapeutic recommendations in selected study settings; however, the availability of an earlier prediction does not necessarily demonstrate that treatment decisions are more accurate or that patient outcomes are improved [41, 43].

AI may support antimicrobial stewardship by aiding in the choice of appropriate antibiotics and decreasing the prescription of unnecessary broad‐spectrum antibiotics. A 2025 comparative meta‐analysis of 27 studies, including approximately 550,000 patients, found that ML models outperformed traditional risk‐scoring systems in sensitivity and negative predictive value, but not in area under the curve, accuracy, specificity or positive predictive value when random‐effects models were applied. The authors also reported substantial heterogeneity, limited external validation and under‐representation of outpatient and pediatric populations. These findings suggest that AI may be most useful for ruling out resistant infections and supporting de‐escalation from broad‐spectrum therapy, although reductions in inappropriate prescribing still need to be demonstrated in prospective implementation studies [41].

Similarly, a systematic review and meta‐analysis of 80 studies on AI in antimicrobial stewardship reported pooled estimates of 72.3% for the area under the curve, 76.9% for sensitivity and 73.8% for specificity. These values indicate moderate rather than high discriminative performance, with considerable heterogeneity across settings and study designs [11]. The broader public‐health antimicrobial stewardship literature describes applications of AI/ML in resistance surveillance, prediction of resistance patterns, outbreak forecasting, antimicrobial prescribing support, and evaluation of stewardship interventions [8].

Clinical decision support integrated into electronic health records utilizes AI to provide recommendations at the POC. These tools analyze sociodemographic characteristics of patients, previous antibiotic usage, laboratory results, and local resistance to a specific antibiotic, dosing, and duration, which may enhance confidence in prescribing appropriate antibiotics [42]. Finally, pharmacists play a key role in interpreting AI‐supported POC results. However, further studies are still being developed considering the role of pharmacists in interpreting AI‐assisted tools in AMR. The literature shows that pharmacists supported by advanced AI analytics can reconsider medication decisions, enhance adherence to stewardship protocols, and improve antibiotic use at the POC. Integration of AI into the pharmacist's practice enables therapy adjustments and supports antimicrobial therapy decisions [44].

Pharmacists may provide an important layer of human oversight by contextualizing AI‐supported outputs, assessing antimicrobial appropriateness, and integrating model recommendations with stewardship principles. Evidence specifically evaluating pharmacist‐mediated implementation of AI‐enabled AMR diagnostics, however, remains limited [8, ⁹]. Evidence from infectious‐disease management in older adults further illustrates the expanding application of AI to diagnosis, prognosis, and therapeutic decision support, while emphasizing the importance of population‐specific validation and clinical interpretability [9].

5. Challenges, Limitations, and Ethical Considerations

Although immensely powerful and theoretically usable in LMICs, current AI models proposed for use in AMR diagnostics, like ML and DL, face challenges. A lack of infrastructural support (i.e., skilled operators, detection systems), which is essential for systemwide or nationwide implementation, has been the most prominent challenge documented so far [45]. Generalizability is another notable challenge that has been tackled in recent research, with proposed models applied to multiple countries in Africa, with accuracy ranging between 45% and 94% across different antibiotics. More broadly, recent evidence syntheses highlight substantial heterogeneity across AI/ML studies, including differences in datasets, pathogens, outcomes, model architectures, and validation strategies [11, 41]. High performance in a retrospective or single‐center dataset therefore does not necessarily indicate equivalent performance in independent hospitals, geographic regions, patient populations, or microbiology laboratories. External validation is essential before routine clinical deployment.

The most currently debatable challenge facing AI involvement in healthcare, generally and in diagnostics specifically, is ethical consideration. This includes fair analysis of algorithms for data, ensuring adequate data privacy, and obtaining consent for data use at a scalable level [46]. Lawmaking is enacted across multiple countries to address equitable care using AI modules. This is taking place in the USA by the Food and Drug Administration, in the UK by the Medicines and Healthcare Products Regulatory Agency, and in Europe by the General Data Protection Regulation [47].

Data quality is considered the basis of building equitable detection models, and ensuring good‐quality datasets with equal numbers of resistant and sensitive colonies to base such models on has remained a persistent challenge in current incorporation [48]. Not only that, but validating these models after development poses another challenge by proxy, which is highly variable and depends on many factors like demographic location and family of antibiotics used [49]. Regulatory approvals have become less of a challenge since 2018, specifically for AI/ML‐enabled tools, which totaled 531, most of which apply to radiology, but 5 of which are specific to use in microbiology [50].

6. Future Directions and Research Gaps

Implementation of AI in the microbiology sector is generally well approached and extremely well researched in the in vitro setting, and some of which is even considered an increasingly hot topic of research in terms of hypothesizing future AMR trends and extrapolating from existing data [51]. Surveillance has also been investigated in the same settings [52]. However, significant limitations remain for each objective of AMR diagnostic modules. For prediction and extrapolation, only single‐reference genome data were used in building the model [53]. For annotating antibiotic resistance genes, application scenarios cannot work on short reads and need time to perform pre‐processing, which makes implementation more difficult [54]. Models designed to target vancomycin intermediate species are also limited, as they were built using only 25 genomes [55], and models predicting carbapenem resistance in A. baumannii and MRSA were developed without any comparison models, highlighting significant comparability bias [56, ^{57, 58, 59, 60, 61, 62, 63}].

7. Conclusion

AI‐driven diagnostic tools represent a promising strategy for bridging the gap between laboratory innovation and bedside application in AMR. By enabling earlier analysis of microbiological, genomic, and clinical information, these technologies may facilitate faster resistance prediction and support targeted antimicrobial decision‐making. However, evidence currently spans clinically evaluated applications, experimental models, and early‐stage technologies, and these levels of evidence should not be considered equivalent. Nevertheless, their long‐term success can be maintained through rigorous clinical validation, harmonized regulatory frameworks, and integration into routine health systems. Equitable implementation, particularly in low‐ and middle‐income settings disproportionately affected by AMR, is critical to avoid widening global health disparities. Sustained investment, capacity building, and cross‐sector collaboration will be essential to ensure scalable, ethical, and lasting global impact.

Author Contributions

Nadine Mugisha: writing – original draft, writing – review and editing, methodology, formal analysis, data curation. Lolita Fawaz: writing – original draft, methodology, formal analysis. Rawan Noureddine: writing – original draft, methodology, formal analysis. Pascal Mathew Okorobe: writing – original draft, methodology, formal analysis. Mohamad Monif Assker: formal analysis, writing – original draft, methodology. Olivier Uwishema: writing – original draft, writing – review and editing, project administration, supervision, conceptualization.

Funding

The authors have nothing to report.

Conflicts of Interest

The authors declare no conflicts of interest.

Data Availability Statement

The authors have nothing to report.