Harnessing Artificial Intelligence to Tackle Antimicrobial Resistance

Foreword

Antimicrobial resistance (AMR) is one of the most significant threats to global health security.

The decade ahead will be critical in developing the necessary resilience to prevent and mitigate the burden of resistant infections, which is estimated to include 39 million deaths worldwide in the next 25 years. A disproportionate number of these will occur in lower- and middle-income countries, and amongst vulnerable populations.

Artificial Intelligence (AI) is already leading to a paradigm shift in healthcare and life science research. The potential to utilise multi-modal data - spanning prevention, diagnostics, surveillance and treatment - has never been so great. However, we are yet to see a significant impact on our shared efforts to address AMR - an area where AI has the potential to help save many lives.

This unrealised potential led us, and our teams at Fleming Initiative and Google DeepMind, to ask one simple question: What do we, as a global community, need to do to leverage AI responsibly and equitably to tackle AMR? We posed this question at a roundtable with global leaders, policymakers, industry executives and research leads, jointly hosted as part of the United Nations fringe events for the high-level meeting on AMR in New York in 2024.

This report outlines the results of these discussions. It highlights the current state of the AI for AMR landscape and offers tangible calls to action to help accelerate progress.

The Fleming Initiative and Google DeepMind are pleased to work collaboratively on this topic. We both believe that no one institution can make this happen, and that siloed working has hindered this field for too long. We hope this report acts as a starting point to bring all stakeholders into the conversation, so they can bring about the innovation needed to change the global trajectory.

Executive Summary

Antimicrobial resistance (AMR) is one of the greatest threats to health worldwide. An increasing number of infections are not treatable with first or second-line antimicrobials. Given AMR is multifaceted in nature, with multiple sectors, geographies and population subgroups variably affected, developing solutions can be complex. These challenges are compounded by the fact that resistance patterns change over time, and therefore, diagnostics, treatments and surveillance systems will need to adapt dynamically. AMR will probably  never be considered “solved”: rather it demands continuous monitoring and management, as it arises from the inherent ability of microbes to evolve and render antimicrobials obsolete.

The increasing power of artificial intelligence (AI) and data-driven technologies in healthcare and scientific research presents a significant opportunity to advance efforts to tackle AMR. However, research on AI for AMR is fragmented, and often concentrated within academic institutions in higher income countries. (1) This report, developed through a critical review of the existing literature, and with input from a breadth of global experts, seeks to identify the necessary changes that will ensure we can better leverage AI when developing solutions to the AMR crisis. It calls upon stakeholders to take action across several areas.

Through these actions, we can encourage collaboration, develop infrastructure, and nurture an ecosystem to advance the field.

The wish list of hypothetical AI applications in AMR is long. To make progress and foster interdisciplinary cooperation we need to isolate the 'machine learnable’ tasks relevant to AMR that could unlock the greatest impact and clearly define them in terms suitable for Machine Learning (ML).

Antimicrobial resistance (AMR) represents a collection of phenomena. It is not defined by a particular pathogen, presentation, geography or resistance mechanism. It can affect multiple interacting epidemiological reservoirs, including humans, animals and the environment – as outlined by the One Health framework. (2) As a result, there are many possible leverage points where AI could be used.

AI is already being used in increasingly diverse ways in healthcare and medical research. Models that can detect anomalies, patterns and deviations from sets of expected outcomes have been in use for almost a decade. (3,4) There are almost 500 FDA-approved software as medical devices using these forms of AI, already embedded into clinical care. (5) Multimodal generative AI models offer new solutions: for instance, by integrating the complex data streams from microscopy, genomics, imaging and clinical practice that are seen in AMR. AI systems can perform optimisation tasks in microbiology and virology labs and simulate experiments, saving resources and informing and guiding in-vitro and in-vivo experiments. (6,7) AI can help interpret the complex biology of microbes, identify the function of microbial proteins and design new reagents to further the study of priority pathogens in academic microbiology labs. For novel drug design, biologists are faced with as many as 10⁶⁰ potential options - AI can help explore this large search area faster than traditional trial-and-error or brute-force approaches. (8) AI tools can also assist with hypothesis generation, helping to focus scientific effort upon questions that are the most viable given current scientific data and knowledge. (9) Additionally, AI has the potential to significantly impact AMR management by analysing patient-specific data, leading to personalised therapy and precision prescribing. (10)

Given this combination of complexity and opportunity, it is important to define the AMR problem landscape. We must critically evaluate where AI, in any of its guises, can be most effectively leveraged.

Some initial guidance on the priority problems within AMR comes from the World Health Organization (WHO). In June 2023, the WHO set out the global research agenda for AMR in human health, where they identified 40 priority research topics for evidence generation to inform AMR policy by 2030, to guide stakeholders including researchers, funders and policymakers. (11) The topics are segmented into prevention, diagnosis, treatment, cross-cutting and resistant tuberculosis (TB). Whilst AI is not mentioned explicitly, in many areas it could play a significant role, and in some instances it is already being deployed. [Appendix].

However, there is no agreement as to which of these 40 areas constitute the highest priority, or where to focus AI-based efforts. There is also a gap in translating these priorities into machine learning questions or tasks and in identifying which AI predictions would most advance the priorities. Unless this changes, AI for AMR research will remain fragmented - risking thinly spread resources, duplication of some efforts, and neglect of others.

Key problem areas in AMR

Many elements of AMR research and management are extremely challenging with current technologies and may require assistance from AI.

1. Fundamental Research – to accelerate our understanding of pathogenesis,  resistance and generate hypotheses

• Current research processes (identifying, classifying, and processing data e.g. genomics) are costly and time consuming
• There is no guarantee of success or downstream uptake
• Our foundational knowledge is not complete – and there remains a breadth of areas to explore

2. Therapeutics – to expand the antimicrobial pipeline and optimise their use

• Drug discovery is costly and time-consuming
• New mechanisms of action are hard to find
• Existing compound libraries are large, so - determining which molecules have undiscovered antimicrobial potential is labour-intensive
• Optimal dosing for certain conditions and population subgroups (e.g. neonates) is not well described
• Prescribing practices are currently not tailored to an individual’s biology

3. Diagnostics – to reduce the current burden of AMR

• Culturing microbes for identification, and subsequent susceptibility testing, is very often too slow to guide initial prescription decisions
• Traditional methods are prone to error due to limitations in sensitivity and specificity
• Diagnostics can often require sophisticated laboratory infrastructure, which is not always available

 4. Threat Surveillance - to maintain a watch over AMR

• Infections can rarely be tracked completely and effectively
• Data from surveillance represents a delayed picture
• Epidemiological monitoring efforts do not comprehensively test all reservoirs

The sorts of problems most amenable to AI are “learnable”: in essence, given the right inputs, one should be able to predict certain outputs. Good problems for AI share three key characteristics:

1. They should address well-defined challenges or tasks that relate to complex problems, processes or systems where many independent variables may interact.
2. They must involve data-rich domains, where the volume, variety, and quality of data allow models to uncover meaningful patterns that traditional approaches cannot.
3. They should have a clear objective function and, critically, coherent evaluation metrics against which the model’s performance can be measured.

Crucially, if we are to prioritise AI for AMR research effectively, we must consider potential impact. Priority should be given to problems where a solution could deliver tangible benefits, for instance accelerating AMR research, developing antimicrobial candidates or saving lives through early interventions. 

By fostering consensus around the most promising opportunities, the AMR and AI research communities and funders will be better able to channel efforts toward key challenges where meaningful progress is achievable.

While early steps toward this prioritisation have begun, they are still emerging and have yet to influence the global research agenda.

CALL TO ACTION: Develop consensus on the highest priority "learnable" problems in AMR that AI could address

The ability to effectively articulate needs, challenges and opportunities across barriers created around complex fields of research will be necessary if experts in AMR and AI are to successfully work together.

By tackling AMR with AI, we are addressing one of the most complex global health challenges using one of the most advanced technologies today. This introduces an unusually diverse range of experts, specialised methodologies, complex datasets and niche parlances.

Navigating the unique intricacies and nuances of both disciplines requires robust interdisciplinary cooperation and knowledge sharing. Crucially, both groups must be able to clearly articulate what they need to succeed in terms that are meaningful to the other. This necessitates developing a shared vocabulary and understanding of technical and practical constraints and possibilities, enabling AMR specialists and AI scientists to effectively communicate and define joint priorities. Research institutions should foster this intersection by creating and supporting dedicated interdisciplinary roles—which might otherwise be neglected—and investing in robust knowledge exchange.

A nascent step towards a more coordinated AMR research ecosystem was taken with the September 2024 United Nations Declaration on AMR, which established an independent panel for evidence for action. (12) However, the panel’s specific responsibilities and governance are still in development. We urge that its remit explicitly include acting as a facilitator between AMR and AI specialists. This will help bridge the diverse sectors involved in AMR, such as human medicine, veterinary medicine, agriculture, and environmental studies. By fostering communication and collaboration across these fields, the panel can help address the unique challenges and harness the opportunities presented by leveraging AI within multisectoral efforts to combat AMR.

CALL TO ACTION: Create a shared language between the stakeholders involved in tackling AMR with AI

Improving the collection, useability and sharing of AMR-relevant data will enable the training of AI models.

Data is a cornerstone of machine learning. The quality, relevance, and representativeness of training data directly impact the utility and bias of AI models.

The field of infectious disease, including AMR, has benefited from a major expansion of data. Molecular microbiological techniques, such as genomic sequencing and mass spectrometry, provide high-dimensional data that can help uncover the cellular processes driving resistance. (13, 14) Electronic medical records (EMRs) offer opportunities to link clinical and microbiologic data, enhancing our understanding of the clinical implications of AMR. (15) Temporal and geospatial metadata enable the tracking of the transmission of resistance. (16)

Efforts are underway to collect and share more AMR data. In November 2024, the Jeddah Commitments at the 4th Global High-Level Ministerial Conference on Antimicrobial Resistance (AMR) outlined support for collecting surveillance data and making it accessible through repositories. (17) Meanwhile, global surveillance networks such as GLASS and regional efforts like EARS-Net and ReLAVRA+ provide frameworks for harmonised data collection. These initiatives have the potential to transform global efforts against AMR by addressing critical gaps.

Data Challenges

However, there are multiple problems with the collection and sharing of AMR-related data. Some crucial data is not being collected; other data is collected but is neither usable nor comparable; and much of the existing data is not being effectively shared. (18-20) As a result, AI cannot yet be used to its full potential in AMR research.

Many crucial types of data are not being collected, at least not systematically. Research into AMR relies on diverse types of data, such as ‘-omic’ data that enable research into the biological mechanisms and determinants of AMR, longitudinal data that tracks how systems and outcomes change over time and disaggregated data for tracing AMR outbreaks as they spread through populations. However, current data collection efforts can be patchy, especially in areas where resources for data collection, storage, and processing are limited. Concerningly, there are few centres for AMR surveillance in Africa, despite recent efforts to scale-up surveillance by the Africa CDC. Countries like Gambia and Mozambique have only one site each where such genomic surveillance is possible. (21) Many other regions also lack robust surveillance capabilities. Without urgent action to fill these gaps, geographic blackspots will persist, undermining global efforts to combat AMR.

Data collection is also often limited and unsuitable for AI. In many contexts, healthcare datasets are still manually entered into multiple, incompatible software systems. This relies on labour-intensive processes that are often prone to human error. Moreover, quantitative data from some sources is kept in analogue formats such as handwritten notes. (22) All these datasets are poorly readable or inaccessible, requiring significant processing (cleaning) before being used for model development. (22,23)

Much of the existing data remains fragmented and inconsistent. Even where systems to capture and store AMR-relevant data exist, they are frequently siloed, limiting accessibility and interoperability. For example, the UK—despite being among the highest-ranking countries for digital maturity—has over 20 different medical record providers and no centralised microbiology database. (24,25) Globally, data collection is often inconsistent, collected at many levels by different agents in varying contexts. The lack of standardisation across collection methods and metadata prevents effective comparison and integration. These problems result in systemic gaps and biases that hinder the development of effective and equitable AI models.

Data sharing is significantly constrained by technical and political barriers, limiting researchers’ ability to access and use relevant data for AMR research. Inadequate digital infrastructure, such as insufficient connectivity and data storage capacity, further constrains the utility of AMR data. (22) Major contributors like the US, China, and Australia do not share full AMR and antimicrobial consumption data with consolidated global systems like GLASS. Much AMR-relevant data, such as pharmacological toxicity data, remains under proprietary control. Likewise, a significant amount of valuable data is confined within microbiology laboratories, disconnected from broader datasets. (26,27)  Despite advances in deep learning approaches for discovering structural classes of antibiotics, the utility of these models would be greater if they could incorporate in vivo or in vitro determinants of antimicrobial activity e.g. serum protein binding. (28)

To address these challenges, it is essential to strategically fill the gaps in the data, improve the usability of collected data, and enable more widespread and equitable sharing of data.

Data Solutions

Considerable data gaps must be filled through new initiatives. These programmes should prioritise the collection of high-quality, high-impact data that links causes and outcomes. Key types of data include:

1. Dense biological data including ‘-omics’, as well as host and pathogen responses
2. Longitudinal data on AMR at pathogen, patient, and population levels
3. Data of many types (including pharmacokinetic/pharmacodynamic, cytotoxicity and resistance mechanism data) from current geographic blackspots

Naturally this new data collection will require investments to strengthen digital infrastructure, such as connectivity and data storage to enable better data collection and analysis. This is likely true in all contexts, but particularly so in LMICs.

Collecting this AMR relevant data is unlikely to prove valuable if the data is not consistent or reflective of real-world challenges. To facilitate the useability of current and yet to be collected data:

1. Clear and simple standards for data collection, data formats and metadata must continue to be defined and communicated ensuring diverse AMR data sources are comparable and integrable. The sensitivity of molecular microbiologic data to the way it is physically processed necessitates full transparency when curating these datasets. Harmonising data collection and more comprehensive metadata practices will help address inconsistencies, enabling datasets to be linked and made universally comprehensible. Ontologies, like those used in the Comprehensive Antibiotic Resistance Database (CARD), should be adopted more broadly to facilitate retrospective integration of AMR-relevant datasets.
2. The data collected must be relevant and reflective of real-world challenges. For example, care must be devoted to the chosen labels. For instance, whilst sensitivity of a microbe to an antimicrobial is frequently reported as a binary outcome, it is more precisely captured by continuously measured minimum inhibitory concentrations. Selection of these outcome labels may require a trade-off of data availability versus best practice.
3. Where possible and appropriate unstructured and analogue data should be processed and digitised to improve readability. There are technologies that can expedite this process, such as natural language processing, optical character recognition, and intelligent character recognition. These technologies require digital infrastructure: connectivity, data storage capacity, and local expertise, so specific additional support may be required particularly for LMICs.
4. Relevant stakeholders should look to provide access to disaggregated data. For example, to fully understand the epidemiological trends affecting AMR, such as its disproportionate effects on certain groups, researchers need access to disaggregated data - anonymised to permit sharing. Ideally it would be possible to identify individuals - subject to compliance with privacy laws, as well as track cases geographically and longitudinally. Unique health identifiers (UHIs) or unique identification numbers (UINs) must be extended to remote populations to achieve this.

Finally, to ensure that worldwide AMR data collection is as impactful as it can be, particularly in the case of developing and training AI/ML models, the data collected must be made accessible. Four steps are crucial for ensuring that AMR-relevant data can be shared and consolidated for those who intend to help make a positive difference:

1. Reduce data silos by encouraging collaboration and fostering trust among stakeholders. Sharing AMR data – whether from private or public entities – requires clear governance frameworks, robust security measures, and recognition of contributors’ efforts. Whilst incentivisation, such as accreditation or prioritised access to integrated datasets, plays a vital role, it is equally important to create environments where collaboration is seen as mutually beneficial and impactful.
2. Use technology to improve the integration of existing data. AI tools can help harmonise existing data sources. Machine learning algorithms, natural language processing, and pattern recognition can help automate the integration of datasets, expediting the development of global data resources to train AI.
3. Open infrastructures must be developed to facilitate equitable data-sharing practices. Efforts should prioritise improving the sharing of outputs as well as original datasets, with mechanisms that account for varying digital capabilities globally. Funding mechanisms should ensure free access to these systems in LMICs, and the value gained from shared systems must be equitably distributed among contributors. An example of this is Vivli’s AMR Register which gathers surveillance data collected by pharmaceutical companies and shares it openly via a single platform.
4. Bespoke solutions for certain types of AMR data sharing, such as federated data platforms used for EMRs in some contexts, should also be considered. Enabling the sharing of critical health data across institutions and regions without raising ownership, privacy or sovereignty concerns, these platforms will allow data to remain decentralised and facilitate collaboration between healthcare systems and countries.

To effectively tackle the complex challenge of strengthening the AMR data landscape, it will be necessary to ensure that all stakeholders are motivated to contribute. This means aligning individual goals with broader global objectives. Effective incentives are critical to this and could range from formal recognition and career benefits to financial rewards or priority access to shared resources. Such incentives will ensure that every participant sees tangible value in their contribution to the collective effort.

As noted above, data privacy is paramount in AI research. This is particularly important in AMR when integrating diverse datasets that could include personal medical data. Initiatives to collect data must ensure compliance with applicable data protection regulations when making data accessible. AI methods must also be robust enough for data of different levels of privacy in diverse regulatory environments. We call for a global conversation regarding the complexities presented by AMR data sharing.

If these proposals are implemented in full, the global response to AMR will be better able to leverage the potential of AI and data-driven technologies. Crucially, robust partnerships across governments, health systems, academia, and private organisations will be required to build equitable, inclusive, and effective AMR data systems.

Data priorities to address

CALL TO ACTION: Reduce missing data and maximise data usability by supporting new AMR data collection efforts

CALL TO ACTION: Create an open data infrastructure that enables and incentivises responsible sharing of AMR relevant data and models

CALL TO ACTION: Drive standardisation of data formats and protocols to aid integration of AMR data from differing geographies, timeframes and modalities

Objective and widely trusted mechanisms for measuring success create focus, enable healthy competition and drive progress.

Benchmarks have been pivotal in supporting the development of AI and encouraging its adoption. The use of benchmarking has been central to driving the uptake of AI in fields as broad as image recognition, speech-to-text, gaming challenges and protein folding. (29)

There are multiple advantages to widely agreed metrics of success for AI, especially in a disparate field such as AMR. At their most fundamental level, robust benchmarks measure the ability of a model to perform a chosen task, a meaningful proxy for real-world usefulness. At a more strategic level, they are instrumental in attracting researchers to work on challenges outside of their core interests; this creates important cross-disciplinary discourse and brings new points of view and approaches to priority problems. The ecosystems and forums that develop around assessments and benchmarks help to drive learning and foster innovation. Likewise, the reproducibility and accessibility of benchmarks are essential for building trust in AI models, which is a strong determinant for their eventual use.

If we are to use AI to successfully address the challenges posed by AMR, we will need carefully developed and consolidated benchmarks. It is important to select tasks that are sufficiently attainable to foster initial momentum and demonstrate early successes, thereby attracting further investment and encouraging wider participation. Yet, these tasks must simultaneously remain sufficiently challenging to stimulate innovation and drive the field forward over the long term. Equally, there must be conceptual clarity behind a benchmark, the selected metrics must be honest and reliable indicators of success that account for the complexity of the challenge. Finally, benchmarks can take time to become established and need to evolve over time. Generally this means benchmarks need to be backed by an organization whose goal is to nurture, maintain and improve the evaluation: naturally, this requires sustained investment.

As yet, there are no international benchmarks or evaluation metrics for AI that pertain to resistant infections. (30) This gap, partly due to the issue of problem prioritisation (Chapter 1), must be addressed. Moving forward, as the most important “machine learnable” AMR relevant problems are identified and clarified, community conveners should look to develop corresponding benchmarks and assessments.

Data collection and access are prerequisites for successful benchmarking (Chapter 3).  Good benchmarking requires large, high-quality datasets that are representative, balanced, and free of bias. (31) Unique challenges also arise with AMR datasets, particularly regarding the type of data employed. Molecular microbiology provides high-dimensionality data, such as genomics and proteomics, which offer the potential to predict resistance-related outcomes. However, this diversity in data types, while offering unique opportunities, also introduces multimodal analytic challenges, such as: integrating data from multiple sources across diagnostics, clinical and demographic data; scalability, where handling large datasets from multiple sources will pose computational and storage challenges, and determining cross-modality between data sources and formulating correct associations. (32,33)

Additionally, AMR patterns differ significantly over time and across regions. Addressing this requires the inclusion of diverse datasets that are temporally and geographically representative to ensure accurate and generalisable findings. (16) Variations in laboratory methodologies—such as antimicrobial susceptibility testing protocols and sequencing platforms—can further complicate benchmarking, leading to inconsistent results. Therefore, standardisation of methods and robust quality control measures are essential to ensure comparability and reliability across datasets.

Benchmarking requires substantial infrastructure investment. AI benchmarking and blind assessments require dedicated software libraries, secure environments, and data storage. (34,35) Creating a representative dataset for AMR research involves combining thousands of bacterial genomes with proteomics data and clinical metadata. This process requires substantial storage capacity, especially for benchmarking and blind assessments. Similarly, performing benchmarking and conducting blind assessments requires investment in adequate computing resources.

Benchmarking should be a catalyst for learning and innovation, accelerating the development of impactful solutions to AMR. The insights generated through these benchmarks should be actively translated into tangible advancements, such as novel therapies, diagnostics, or preventative strategies.  This requires a robust and coordinated ecosystem encompassing both tools and infrastructure. Tools, including AI algorithms, data analysis frameworks, and simulation models, should be designed for interoperability and reproducibility. Similarly, infrastructure, such as data repositories, computational resources, and collaborative platforms, should facilitate data sharing and integration across benchmarking initiatives. To maximize the impact of benchmarking, dedicated communities, platforms, and forums focused on AMR are essential for fostering collaboration, knowledge dissemination, and the rapid translation of research findings into practical applications

The way AI for AMR benchmarking is developed now will define AI for AMR research for years to come. Inevitably, decisions about problem selection, data types and outcome selection are built into benchmarking systems: those decisions will shape our future response to AMR (Chapters 1 and 3). Consequently, benchmarks must be developed by building consensus among domain experts from all key AMR-relevant disciplines and from a diverse range of backgrounds.

CALL TO ACTION: Develop and agree robust evaluation metrics and benchmarks to demonstrate the effectiveness of AI systems developed for AMR problems

CALL TO ACTION: Nurture the emergent evaluation ecosystem by investing in the necessary resources, infrastructure, blind assessments and forums, that catalyse innovation and knowledge sharing

Case Study

Critical Assessment of Structure Prediction (CASP)

Predicting the three-dimensional structures of proteins from their amino acid sequences was an extraordinary technical challenge. It attracted the interest of academics and researchers in biochemistry, due to both the task’s complexity and the enormous potential value of a solution.

A biennial challenge called Critical Assessment of Structure Prediction was launched in 1994. During CASP, research groups submit predicted protein structures, which are compared against experimentally determined structures that have not yet been made public.

Initially, the niche nature of the challenge led to limited improvements in model accuracy. However, with the engagement of a wider community and the introduction of new methods like deep learning, the 12^th and 13^th iterations saw a stepwise increase in model accuracy.

AlphaFold, developed by Google DeepMind, won the CASP 13 and CASP 14 competitions. It now offers the ability to accurately predict the structures of all known proteins, plus complexes formed by protein interactions with RNA, DNA, ligands and ions. AlphaFold is helping to drive a wave of innovation in the health and life sciences.

Similarly rigorous benchmarking of AI for AMR could also drive long-term improvements in models and predictions. The CASP story shows that such progress is not necessarily immediate but requires long-term investment in the benchmarking process.

Diverse human expertise is crucial for the responsible development and use of AI. Investing in talent, especially in LMICs, will boost capacity and inclusion.

Evidence suggests the global AMR research and development workforce is diminishing. Estimates from the AMR Industry Alliance suggest there are only 3000 clinical AMR researchers active worldwide; that is 15 times fewer than cancer, and 1.5 times fewer than HIV/AIDs. (36) There are also declining numbers of infectious disease specialists, even in high-income countries. A 2021 workforce survey in the UK showed 17.5% of all funded full-time infection consultant posts were vacant. In the US, the IDSA have reported that 80% of US counties have no infectious disease specialists. (37, 38) This so-called ‘brain-drain’ from both research and clinical activity is largely due to a lack of dependable long-term funding, even in high-income countries. Without a resilient, skilled workforce, our capacity to develop and implement new solutions to address AMR is threatened.

In contrast, the global AI talent pool has expanded significantly over the past decade, with the health sector a particular focus for recruitment. However, the distribution of AI proficiencies is highly concentrated in a limited number of regions, with the US the primary destination for experts and institutions. 70% of all AI talent is concentrated in only 5 countries, with 50% of the “top” European AI talent produced in 3 countries. (39) Meanwhile, Latin America, the Caribbean and Sub-Saharan Africa produce the lowest volumes of peer-reviewed AI research publications. (29)

The highly uneven distribution of AI expertise and the relatively small and diminishing global population of AMR researchers has significant implications for work at the intersection of AI and AMR. Few institutions possess the expertise in both areas needed to progress. Furthermore, the acute under-representation of some regions makes it likely that biases or a lack of contextual understanding will exist - in which case, AI development will further perpetuate inequities. This was recently demonstrated by a COVID-19 case study, in which datasets sourced from distinct healthcare settings across the globe created biases in algorithms. However, the study also showed that implementing algorithmic-level bias mitigation significantly improved outcome fairness between LMICs and HICs. (40)

However, it is worth noting that several LMICs have demonstrated significant strength and progress in leveraging AI in healthcare. Countries such as India and Brazil have developed innovative uses of AI to address local challenges, including in the use of telehealth in remote areas. (41,42) These successes in health-related AI highlight the untapped potential of LMICs to contribute meaningfully to AI for AMR. Supporting and expanding this local expertise could enable the development of AI-driven AMR solutions that are both globally relevant and contextually appropriate.

Training programmes at the intersection of AI and AMR must form part of this solution. These programmes should focus on building interdisciplinary expertise by equipping AI experts with sufficient AMR knowledge to apply their skills effectively, whilst also supporting AMR professionals to develop AI competencies.

However, these initiatives alone are insufficient without stable career pathways underpinned by long-term funding for currently neglected interdisciplinary roles. Governments, international organisations, and private entities must collaborate to embed sustainable funding and support mechanisms to create pathways for training and careers at the intersection of AI and AMR. These programmes should:

• Span multiple regions to ensure inclusion, especially in underrepresented areas. Training efforts must prioritise global representation, actively involving under-resourced communities and regions to address disparities in education and career development opportunities.
• Engage public and private sectors to harness diverse expertise. Partnerships with pharmaceutical companies, technology firms, healthcare providers, and public institutions are crucial. Such collaborations can involve co-developing curricula, establishing mentorship initiatives, and creating funding mechanisms. Dedicated task forces, public-private grants, and knowledge-sharing platforms can facilitate these efforts, ensuring solutions are practical and aligned with real-world needs.
• Support researchers at all career stages. Initiatives such as AMR-focused early-career fellowships and targeted funding opportunities can help build a sustainable interdisciplinary workforce while providing structured career progression pathways.

The benefits of such integrated programmes include creating a sustainable talent pipeline, enhancing interdisciplinary collaboration, and ensuring the practical adoption of AI solutions. Only by addressing both the immediate need for skills development and the long-term sustainability of AMR-related careers can we build a workforce capable of developing equitable, innovative AI solutions for AMR.

Case Study

Capacity Accelerator Network (CAN) Data Science Fellowship (43)

The CAN fellowship supports early career researchers from across Africa who are undertaking data science projects at the intersection of health and climate in Africa, especially focused on land degradation.

Funded by the Wellcome Trust and Group on Earth Observations - Land Degradation Neutrality (GEO-LDN), it is an example of an integrated programme to build knowledge and capacity to develop data-driven technologies that can address an area of significant global concern.

In AMR such opportunities are currently lacking. With this in mind, the Fleming Initiative and Google DeepMind are proud to announce the establishment of a new, fully-funded three-year postdoctoral Academic Fellowship that will support AI for AMR research.

Building on the established Google DeepMind Academic Fellowship funding stream, this new fellowship has been created to provide an opportunity for an early career researcher to develop the interdisciplinary capabilities needed to realise the potential of AI for AMR.

The Academic Fellow will be independently appointed by the Fleming Initiative, and free to research any topic within AI and AMR. They will also be offered mentoring from senior Google DeepMind researchers. The chosen Fellow will also play an important role in fostering inclusive and equitable approaches to training and skill-building for the current and next generation of AI and AMR researchers.

Recruitment for the Google DeepMind Academic Fellowship in AMR and AI will be announced by Fleming Initiative later in 2025.

CALL TO ACTION: Develop the necessary interdisciplinary capabilities and skills needed to unlock AI solutions for use in AMR

To enable more locally effective AI for AMR, we must reduce inequities in access to AI, including both development and deployment.

AMR can affect everyone, but its impact is not equally distributed. It disproportionately affects vulnerable populations, particularly those in low- and middle-income countries (LMICs). These regions often face significant challenges, including inadequate healthcare infrastructure, limited access to diagnostics, and overburdened health systems - all of which exacerbate the impact of AMR.

AI has great potential to reduce these inequities in AMR – but only if AI systems are truly representative and effective across various global contexts. To achieve this, AI infrastructure must be developed and deployed equitably, particularly in contexts that currently cannot fully harness AI. These regions must be able to meaningfully contribute to the development of AI, by conducting AI-in-AMR research and providing the vital data required for algorithms to perform well in local populations and health systems.

Innovative uses of AI are progressing in LMICs. For instance, AntiMicro.ai, an AI-powered application to predict antibacterial and antifungal susceptibility was recently developed and deployed by researchers in Kenya. (44) The tool re-uses Pfizer’s ATLAS dataset, which was made available through the NGO Vivli’s AMR Data Challenge. AntiMicro.ai underscores the capacity of LMICs to contribute substantially to global AI solutions, particularly when resources and expertise are directed toward locally relevant challenges. (45)

However, there remain many barriers to developing and deploying AI for AMR in LMICs. In some areas there remains a lack of computer and connectivity infrastructure including poor internet and mobile coverage; lack of data storage and processing capacity; and little access to computers capable of training or running AI models. (46) These challenges are exacerbated by limited resources, by a lack of an AI-literate workforce, and challenges with data collection in AI-appropriate formats (Chapter 3). To note, the researchers that developed AntiMicro.ai faced problems of data-missingness and geographical skew. (44)

Significant support is therefore required to enable widespread access to AI and address the aforementioned barriers. This could take several forms:

• Local data centres and high-performance computing clusters capable of running diverse AI models could be established in LMICs. These would enable more diverse LMIC-based AI for AMR research and help with implementing models within local or regional health systems. These centres would enable faster and more localised data processing, as well as bespoke programs tailored to LMIC disease burden and infection prevention and control (IPC) contexts. (47) However, the energy and infrastructure requirements may be considerable.
• Cloud computing solutions could eliminate the requirement for local physical data centres. However, they require reliable connectivity, so would still entail physical infrastructure development in many regions. Also, owing to data sensitivities and local laws, it may not always be possible to export certain types of health data - meaning cloud solutions may not be appropriate in all circumstances.
• Innovative approaches like Edge AI (where AI programs run offline and natively on personal devices) and Distributed AI (where AI computing tasks are spread across multiple devices to balance computational power) may offer a third way. (48,49) However, these approaches could present challenges of their own, including the requirement for large upfront investments into models trained on traditional computing clusters that are capable of running in low-resource settings.

Developing AI models for LMICs will require appropriate training data from these regions (Chapter 3). Developers of AI for AMR should look to prioritise models that can use data in forms that are consistently accessible in LMICs. For instance, in some settings and for some problems, AI models may ultimately be more practically useful, and therefore impactful, if trained on clinical or microbiological images rather than -omic data.

Case study

Antibiogo (50)

Antibiotic susceptibility tests (ASTs) are the current gold standard for determining which antimicrobials bacteria are susceptible to, and conversely which they are resistant to. Test results determine the best antibiotic options for the treatment of bacterial infection.

The most rudimentary AST is disc diffusion, where bacterial isolates from patients are plated on agar plates interspersed with antibiotics. The larger the zone of inhibition (where the bacteria do not grow due to action of the antibiotic) - the more the antibiotic can inhibit the bacteria. Conversely, a small or non-existent zone of inhibition might suggest bacterial resistance to the antibiotic.

Antibiogo is an easy-to-implement app-based tool which uses machine learning to measure the inhibition zone diameter on agar plates using a phone camera. The app uses image processing and machine learning to produce an interpretation of the AST and suggests the next steps. It also scales up expert intelligence by defining the best course of treatment for clinicians, lab technicians and nurses.

Running locally on devices, Antibiogo eliminates the need for a reliant internet connection and is currently free to use, allowing its potential widespread use in the field when fully released.

Deploying AI for AMR into LMIC healthcare systems poses its own challenges. Successful healthcare AI deployment requires a strong digital public infrastructure (DPI), defined as “a set of shared digital systems that are secure and interoperable, built on open technologies, to deliver equitable access to public and/or private services at a societal scale". A strong DPI should facilitate the development of healthcare systems ready to safely deploy AI systems for AMR management.

The G7 underscored the importance of DPI during its 2023 summit, through initiatives such as the Partnership for Global Infrastructure and Investment (PGII). This partnership is committed to funding projects that foster secure and interoperable digital systems globally, including investments in health-focused digital solutions. (51) Notably, the G7 supported collaborations such as the Global Digital Health Initiative, aimed at strengthening health systems across borders and ensuring equitable access to AI-ready infrastructure. (52) These efforts have led to tangible international impacts, for example partnerships to accelerate AI integration into diagnostics and treatment, and investments targeting under-resourced regions, ensuring global participation in the fight against AMR. By aligning with these global initiatives, countries can support the safe deployment of AI-driven solutions for AMR management.

Supporting widespread access to AI is essential for empowering LMICs to tackle AMR effectively. Recognising and building on their existing strengths will ensure that the global fight against AMR is inclusive, equitable, and grounded in diverse contexts and perspectives.

CALL TO ACTION: Support access to AI in LMICs to enable the equitable development and deployment of AI for AMR globally

To achieve the maximum positive impact, AI for AMR must be used responsibly and offer benefits to users around the world.

As AMR disproportionately impacts vulnerable populations – such as the young, elderly, and those in low-resource settings – it is critical to ensure that AI models developed for AMR are both responsible and equitable . By adopting a ‘responsible-by-design’ approach, we can harness AI’s potential to combat AMR while fostering trust, collaboration, and equitable outcomes worldwide.

There are several frameworks that support a responsible approach to developing and deploying AI. For instance, the WHO outlines six principles when designing, developing, and deploying AI for health:

1. Protect autonomy
2. Promote human well-being, human safety, and the public interest
3. Ensure transparency, explainability, and intelligibility
4. Foster responsibility and accountability
5. Ensure inclusiveness and equity
6. Promote AI that is responsive and sustainable.

AMR, and infectious diseases more broadly, present unique challenges. These include developing global data resources whilst preserving privacy; balancing the careful use of antibiotics with ensuring the best possible care for individual patients; and accounting for the evolving nature of antimicrobial resistance.

At present there is no consensus on the most important principles that should be applied to AI for AMR. In the table below, we outline some potential considerations that might be applied to AI for AMR, using the WHO principles as a template. We call for a wider discussion on how to apply ethical principles to AI for AMR.

A responsible approach to AI also requires understanding and mitigating potential misuse of tools. Moving forward, it will continue to be important to closely monitor the impact and risks of new models.

Applying a responsible-by-design approach will help facilitate trust in AI systems used for AMR. It will also improve global efforts on AMR in general by fostering collaboration. We therefore call for the global adoption of strong ethical frameworks in AI for AMR research.

CALL TO ACTION: Adopt a “responsible by design” approach to AI for AMR to ensure beneficial and equitable outcomes

AMR is one of the most critical global challenges we face, whilst AI is a technology that has the potential to transform both healthcare, and the scientific research that underpins it. At the convergence of AI and AMR is a diverse range of experts, specialised methodologies, multi-dimensional datasets and niche parlances. The complexity of this landscape brings about scientific, political, cultural and engagement barriers, all of which can hinder progress.

It will take an interdisciplinary, multi-faceted approach to overcome many of these challenges. 2024 saw a welcome increase in the levels of interest in AMR, with the United Nations General Assembly and 4th Ministerial meeting providing opportunities to foster greater international cooperation. However, for the potential of AI for AMR to be realised, these efforts must be sustained using coordinated and phased action.

Here we sketch out a roadmap for the next five years, as a starting point for consensus-building and planning. In the next year, efforts must focus on fostering collaboration through global workshops, defining priority questions and initial benchmarks, developing a shared language with communication toolkits, and mapping critical data gaps. Within two years, this foundation must evolve into validated data standards, piloted data-sharing platforms, and the incorporation of AI modules into infectious disease training pathways. By the five-year mark, these actions should culminate in globally embedded interdisciplinary curricula, expanded AI infrastructure in LMICs, routine benchmarking of AI systems for instance through global competitions, and the widespread adoption of ethical frameworks to guide responsible deployment.

While this roadmap provides a structured pathway, it remains non-exhaustive and questions remain on financial constraints, political dynamics, technological readiness and who is accountable for tracking success. Global collaboration across governments, academia, industry, and health organisations will be essential to translate these milestones into tangible progress. Accountability mechanisms may also be needed to ensure timely progress. Through sustained commitment, the global community can leverage AI to combat AMR effectively and equitably, setting a precedent for addressing other complex health challenges.

By implementing the recommendations of this report, future efforts to use AI for AMR can be made practicable, responsible and more effectively coordinated. Successful deployment of AI for AMR will help to maximise global benefit, at a time when solutions are desperately needed.

Acknowledgements

This authorship group for this report was co-chaired by Lord Ara Darzi (Executive Chair of the Fleming Initiative) and Anna Koivuniemi (Head of Google DeepMind Impact Accelerator).

The report was co-authored by Amish Acharya, Simon Dryden, Pushmeet Kohli, Jack Mason, Eliseo Papa, Anant Pratap Singh and Lakshya Soni.

The authors would like to thank all contributors from Google DeepMind, the Fleming Initiative and the Institute of Global Health Innovation for their input, review and assistance to create this report. In particular they would like to thank Professor Alison Holmes, Conor Griffin,  Agata Laydon and Juan Mateos-Garcia.

In addition the authors would like to acknowledge the role of all roundtable participants at the United Nations meeting. At time of publication, the following participants had formally consented to being named:

Dr Lena Afeyan, Specialist Advisor, Fleming Initiative

Professor Hutan Ashrafian, Chief Scientific Officer, Pre-emptive Medicine & Health, Flagship Pioneering

Dr Jennifer Cohn, Director, Global Access, GARDP

Dame Sally Davies, UK Envoy on AMR

Dr Cecilia Ferreyra, Director AMR Programme, FIND

Dame Vivian Hunt, Chief Innovation Officer, Optum

Dr Daudi Jjingo, Director, African Center of Excellence in Bioinformatics & Data-Intensive Sciences

Dr Alain Labrique,  Director of Digital Health and Innovation, Science Division, WHO

Professor Branwen Morgan, Minimising AMR Mission Lead, CSIRO

Dr Naveen Rao, Senior Vice President of the Health Initiative at The Rockefeller Foundation

Appendix 

Selected AI examples aligned to WHO Global Research Agenda Priority topics

Status refers to whether the example is theorised, currently confined to research, or whether it is in the process of deployment for clinical use.

References

(1) Coque TM, Cantón R, Pérez-Cobas AE, Fernández-de-Bobadilla MD, Baquero F. Antimicrobial Resistance in the Global Health Network: Known Unknowns and Challenges for Efficient Responses in the 21st Century. Microorganisms. 2023; 11 (4): 1050. 10.3390/microorganisms11041050.

(2) World Health Organization. One Health. https://www.who.int/health-topics/one-health#tab=tab_1 [Accessed 3rd Jan 2025].

(3) Wong W, Moore A, Cooper G, Wagner M. Wsare: What’s strange about recent events? Journal of Urban Health : Bulletin of the New York Academy of Medicine. 2003; 80 (Suppl 1): i66–i75. 10.1007/PL00022317.

(4) Carter JH. The Immune System as a Model for Pattern Recognition and Classification. Journal of the American Medical Informatics Association. 2000; 7 (1): 28–41. 10.1136/jamia.2000.0070028.

(5) U.S. Food & Drug Administration. Artificial Intelligence and Machine Learning (AI/ML)-Enabled Medical Devices. https://www.fda.gov/medical-devices/software-medical-device-samd/artificial-intelligence-and-machine-learning-aiml-enabled-medical-devices [Accessed Jan 3, 2025].

(6) Apte P. Accelerating research with AI-assisted experiments. MIT Department of Materials Science and Engineering. -12-20T15:19:31+00:00. 2023. https://dmse.mit.edu/news/accelerating-research-with-ai-assisted-experiments/ [Accessed Jan 9, 2025].

(7) FormBio. Augmenting Biological Lab Experimentation With AI. https://www.formbio.com/blog/biological-lab-ai-applications [Accessed Jan 9, 2025].

(8) Conor Griffin, Don Wallace, Juan Mateos-Garcia, Hanna Schieve, Pushmeet Kohli. S. A new golden age of discovery. https://www.aipolicyperspectives.com/p/a-new-golden-age-of-discovery [Accessed Jan 9, 2025].

(9) Hutson M. Hypotheses devised by AI could find ‘blind spots’ in research. Nature. 2023; 10.1038/d41586-023-03596-0.

(10) Branda F, Scarpa F. Implications of Artificial Intelligence in Addressing Antimicrobial Resistance: Innovations, Global Challenges, and Healthcare’s Future. Antibiotics. 2024; 13 (6): 502. 10.3390/antibiotics13060502.

(11) World Health Organization. Global research agenda for antimicrobial resistance in human health. World Health Organization; 2024.

(12) Anonymous Political Declaration of the High-level Meeting on Antimicrobial Resistance. 2024.

(13) Sauerborn E, Corredor NC, Reska T, Perlas A, Vargas da Fonseca Atum S, Goldman N, et al. Detection of hidden antibiotic resistance through real-time genomics. Nature Communications. 2024; 15 (1): 5494. 10.1038/s41467-024-49851-4.

(14) Kaprou GD, Bergšpica I, Alexa EA, Alvarez-Ordóñez A, Prieto M. Rapid Methods for Antimicrobial Resistance Diagnostics. Antibiotics. 2021; 10 (2): 209. 10.3390/antibiotics10020209.

(15) Blechman SE, Wright ES. Applications of Machine Learning on Electronic Health Record Data to Combat Antibiotic Resistance. The Journal of Infectious Diseases. 2024; 230 (5): 1073–1082. 10.1093/infdis/jiae348.

(16) Fu X, Chen C, Dong Y, Vullikanti A, Klein E, Madden G, et al. Spatial-Temporal Networks for Antibiogram Pattern Prediction. ArXiv. 2023; arXiv:2305.01761v1. https://pubmed.ncbi.nlm.nih.gov/37205267/. .

(17) World Health Organization. Global ministers and partners pledge action with new Jeddah Commitments on AMR. https://www.who.int/news/item/21-11-2024-global-ministers-and-partners-pledge-action-with-new-jeddah-commitments-on-amr [Accessed 9th Jan 2025].

(18) Iskandar K, Molinier L, Hallit S, Sartelli M, Hardcastle TC, Haque M, et al. Surveillance of antimicrobial resistance in low- and middle-income countries: a scattered picture. Antimicrobial Resistance & Infection Control. 2021; 10 (1): 63. 10.1186/s13756-021-00931-w.

(19) Robillard DW, Sundermann AJ, Raux BR, Prinzi AM. Navigating the network: a narrative overview of AMR surveillance and data flow in the United States. Antimicrobial Stewardship & Healthcare Epidemiology : ASHE. 2024; 4 (1): e55. 10.1017/ash.2024.64.

(20) Kiggundu R, Lusaya E, Seni J, Waswa JP, Kakooza F, Tjipura D, et al. Identifying and addressing challenges to antimicrobial use surveillance in the human health sector in low- and middle-income countries: experiences and lessons learned from Tanzania and Uganda. Antimicrobial Resistance & Infection Control. 2023; 12 (1): 9. 10.1186/s13756-023-01213-3.

(21) Okolie OJ, Igwe U, Ismail SU, Ighodalo UL, Adukwu EC. Systematic review of surveillance systems for AMR in Africa. Journal of Antimicrobial Chemotherapy. 2022; 78 (1): 31–51. 10.1093/jac/dkac342.

(22) Gandra S, Alvarez-Uria G, Turner P, Joshi J, Limmathurotsakul D, van Doorn HR. Antimicrobial Resistance Surveillance in Low- and Middle-Income Countries: Progress and Challenges in Eight South Asian and Southeast Asian Countries. Clinical Microbiology Reviews. 2020; 33 (3): 48. 10.1128/CMR.00048-19.

(23) Siyam A, Ir P, York D, Antwi J, Amponsah F, Rambique O, et al. The burden of recording and reporting health data in primary health care facilities in five low- and lower-middle income countries. BMC Health Services Research. 2021; 21 (1): 691. 10.1186/s12913-021-06652-5.

(24) Warren LR, Clarke J, Arora S, Darzi A. Improving data sharing between acute hospitals in England: an overview of health record system distribution and retrospective observational analysis of inter-hospital transitions of care. BMJ open. 2019; 9 (12): e031637. 10.1136/bmjopen-2019-031637.

(25) Clews G. Electronic patient records being used in 90% of hospital trusts, says NHS England. https://pharmaceutical-journal.com/article/news/electronic-patient-records-being-used-in-90-of-hospital-trusts-says-nhs-england [Accessed Dec 20, 2024].

(26) Access to Medicine Foundation. 2021 Antimicrobial Resistance Benchmark | Access to Medicine. 2021. https://accesstomedicinefoundation.org/resource/2021-antimicrobial-resistance-benchmark. .

(27) Okeke IN1, Feasey N2, Parkhill J3, Turner P4, Limmathurotsakul D5, Georgiou P6, et al. Leapfrogging laboratories: the promise and pitfalls of high-tech solutions for antimicrobial resistance surveillance in low-income settings. 2020; e003622. 10.1136/bmjgh-2020-003622.

(28) Wong F, Zheng EJ, Valeri JA, Donghia NM, Anahtar MN, Omori S, et al. Discovery of a structural class of antibiotics with explainable deep learning. Nature. 2024; 626 (7997): 177–185. 10.1038/s41586-023-06887-8.

(29) Stanford AI Index Report. 2021; https://aiindex.stanford.edu/wp-content/uploads/2021/03/2021-AI-Index-Report-_Chapter-1.pdf. .

(30) Ali T, Ahmed S, Aslam M. Artificial Intelligence for Antimicrobial Resistance Prediction: Challenges and Opportunities towards Practical Implementation. Antibiotics. 2023; 12 (3): 523. 10.3390/antibiotics12030523.

(31) Sourlos N, Vliegenthart R, Santinha J, Klontzas ME, Cuocolo R, Huisman M, et al. Recommendations for the creation of benchmark datasets for reproducible artificial intelligence in radiology. Insights into Imaging. 2024; 15 248. 10.1186/s13244-024-01833-2.

(32) Fortuin S, Soares NC. The Integration of Proteomics and Metabolomics Data Paving the Way for a Better Understanding of the Mechanisms Underlying Microbial Acquired Drug Resistance. Frontiers in Medicine. 2022; 9 849838. 10.3389/fmed.2022.849838.

(33) Visonà G, Duroux D, Miranda L, Sükei E, Li Y, Borgwardt K, et al. Multimodal learning in clinical proteomics: enhancing antimicrobial resistance prediction models with chemical information. Bioinformatics. 2023; 39 (12): btad717. 10.1093/bioinformatics/btad717.

(34) Aldoseri A, Al-Khalifa KN, Hamouda AM. Re-Thinking Data Strategy and Integration for Artificial Intelligence: Concepts, Opportunities, and Challenges. Applied Sciences. 2023; 13 (12): 7082. 10.3390/app13127082.

(35) Le DQ, Nguyen TT, Nguyen CH, Ho TH, Vo NS, Nguyen T, et al. AMRomics: a scalable workflow to analyze large microbial genome collections. BMC Genomics. 2024; 25 709. 10.1186/s12864-024-10620-8.

(36) Leaving the Lab: Tracking the Decline in AMR R&D Professionals. AMR Industry Alliance. https://www.amrindustryalliance.org/mediaroom/leaving-the-lab-tracking-the-decline-in-amr-rd-professionals/ [Accessed Dec 20, 2024].

(37) Royal College of Pathologists, British Infection Association. The infection sciences workforce: challenges and solutions. 2023. https://www.rcpath.org/static/27a069ea-3ff6-46bb-b691c26c58a5f16f/The-infection-sciences-workforce-report-Final.pdf. .

(38) Infectious Diseases Society of America, HIV Medicine Association. Grow and Diversify the Infectious Diseases Workforce to Promote Health Equity. 2024.

(39) Marcopolo. The Global AI Talent Tracker 2.0. https://macropolo.org/interactive/digital-projects/the-global-ai-talent-tracker/ [Accessed Jan 3, 2025].

(40) Yang J, Clifton L, Dung NT, Phong NT, Yen LM, Thy DBX, et al. Mitigating machine learning bias between high income and low–middle income countries for enhanced model fairness and generalizability. Scientific Reports. 2024; 14 (1): 13318. 10.1038/s41598-024-64210-5.

(41) Morales HMP, Guedes M, Silva JS, Massuda A. COVID-19 in Brazil-Preliminary Analysis of Response Supported by Artificial Intelligence in Municipalities. Frontiers in Digital Health. 2021; 3 648585. 10.3389/fdgth.2021.648585.

(42) Gore MN, Olawade DB. Harnessing AI for public health: India's roadmap. Frontiers in Public Health. 2024; 12 1417568. 10.3389/fpubh.2024.1417568.

(43) Mwagiru A. Data science for health and climate in Africa: Meet the CAN Fellows. https://www.data4sdgs.org/blog/data-science-health-and-climate-africa-meet-can-fellows [Accessed Jan 11, 2025].

(44) Mutisya F, Kanguha R. AntiMicro.ai: An Artificial Intelligence powered web application for predicting antibacterial/antifungal susceptibility and constructing personalized machine learning models. Wellcome open research. 2024; 9 273. 10.12688/wellcomeopenres.21281.1.

(45) Aliffi K. Breaking News – Vivli announces the 2024 AMR Surveillance Data Challenge, funded by GARDP, Paratek, Pfizer and Vivli, to Initiate June 17th. Vivli AMR. -06-04T13:58:08+00:00. 2024. https://amr.vivli.org/breaking-news-vivli-announces-the-2024-amr-surveillance-data-challenge/ [Accessed Jan 11, 2025].

(46) Khan MS, Umer H, Faruqe F. Artificial intelligence for low income countries. Humanities and Social Sciences Communications. 2024; 11 (1): 1–13. 10.1057/s41599-024-03947-w.

(47) Ciecierski-Holmes T, Singh R, Axt M, Brenner S, Barteit S. Artificial intelligence for strengthening healthcare systems in low- and middle-income countries: a systematic scoping review. npj Digital Medicine. 2022; 5 (1): 1–13. 10.1038/s41746-022-00700-y.

(48) IBM. What Is Edge AI? | IBM. https://www.ibm.com/topics/edge-ai [Accessed Jan 4, 2025].

(49) Chong W, Desai N, Govindjee A, Wang X, Witherspoon S. What is Distributed AI? https://developer.ibm.com/what-is-distributed-ai [Accessed Jan 4, 2025].

(50) Medecins Sans Frontieres. Antibiogo. https://www.antibiogo.org/ [Accessed Jan 4, 2025].

(51) House TW. FACT SHEET: Partnership for Global Infrastructure and Investment at the G7 Summit. https://www.whitehouse.gov/briefing-room/statements-releases/2023/05/20/fact-sheet-partnership-for-global-infrastructure-and-investment-at-the-g7-summit/ [Accessed Jan 4, 2025].

(52) G7 Health Ministers' Meeting 2024. G7 Policy brief on Artificial Intelligence opportunities and challenges (2024). https://g7g20-documents.org/database/document/2024-g7-italy-ministerial-meetings-health-ministers-miscellaneous-g7-policy-brief-on-artificial-intelligence-opportunities-and-challenges [Accessed Jan 4, 2025].

(53) Bhattacharyya RP, Bandyopadhyay N, Ma P, Son SS, Liu J, He LL, et al. Simultaneous detection of genotype and phenotype enables rapid and accurate antibiotic susceptibility determination. Nature Medicine. 2019; 25 (12): 1858–1864. 10.1038/s41591-019-0650-9.

(54) Zagajewski A, Turner P, Feehily C, El Sayyed H, Andersson M, Barrett L, et al. Deep learning and single-cell phenotyping for rapid antimicrobial susceptibility detection in Escherichia coli. Communications Biology. 2023; 6 (1): 1–12. 10.1038/s42003-023-05524-4.

(55) Boyce JM, Laughman JA, Ader MH, Wagner PT, Parker AE, Arbogast JW. Impact of an automated hand hygiene monitoring system and additional promotional activities on hand hygiene performance rates and healthcare-associated infections. Infection Control & Hospital Epidemiology. 2019; 40 (7): 741–747. 10.1017/ice.2019.77.

(56) Senescau A, Kempowsky T, Bernard E, Messier S, Besse P, Fabre R, et al. Innovative DendrisChips® Technology for a Syndromic Approach of In Vitro Diagnosis: Application to the Respiratory Infectious Diseases. Diagnostics. 2018; 8 (4): 77. 10.3390/diagnostics8040077.

(57) Lee S, Shin J, Kim HS, Lee MJ, Yoon JM, Lee S, et al. Hybrid Method Incorporating a Rule-Based Approach and Deep Learning for Prescription Error Prediction. Drug Safety. 2022; 45 (1): 27–35. 10.1007/s40264-021-01123-6.

(58) Beaudoin M, Kabanza F, Nault V, Valiquette L. Evaluation of a machine learning capability for a clinical decision support system to enhance antimicrobial stewardship programs. Artificial Intelligence in Medicine. 2016; 68 29–36. 10.1016/j.artmed.2016.02.001.

(59) AI Diagnostics | AI-Powered TB Screening. https://www.aidiagnostics.health/ [Accessed Dec 17, 2024].

(60) USAID, Wadhwani AI, Central TB Division, Ministry of Health & Family Welfare. AI solutions for TB Elimination: Defining the scope of AI solutions within NTEP. 2021; https://www.wadhwaniai.org/wp-content/uploads/2021/10/1_NTEP_reportV11_020721.pdf. .

(61) Horwitz MA, Clemens DL, Lee B. AI-Enabled Parabolic Response Surface Approach Identifies Ultra Short-Course Near-Universal TB Drug Regimens. Advanced Therapeutics. 2019; 1900086. 10.1002/adtp.201900086.

(62) Kant R, Khan MS, Chopra M, Saluja D. Artificial intelligence-driven reverse vaccinology for Neisseria gonorrhoeae vaccine: Prioritizing epitope-based candidates. Frontiers in Molecular Biosciences. 2024; 11 10.3389/fmolb.2024.1442158.