Impact of artificial intelligence on AMR: Common themes and challenges

Researchers gather to explore visions, themes & challenges around the use of Artificial Intelligence (AI) to tackle antimicrobial resistance (AMR).

On February 9^th, over 55 in-person and 60 online attendees joined a collaborative symposium to explore the impact of AI in research and practice on AMR. The event, co-organized by Imperial College London’s Institute of Infection, the AI Network, the Centre of Antimicrobial Optimisation, and the NIHR Health Protection Research Unit for Healthcare-Associated Infections and Antimicrobial Resistance, showcased the breadth of research being undertaken across Imperial and brought researchers together from different disciplines. In sessions focused on different parts of the translational pipeline, examples were presented of how AI models are being used for AMR prediction and clinical decision support. Future opportunities and barriers were also explored.

Vision for AI models in healthcare: Common themes and challenges

A number of common themes arose, including the quality and completeness of data used as input to AI algorithms. Sessions highlighted the importance of using a range of data methods to account for the inherent diversity of information relating to AMR. Ensuring appropriate implementation of these into clinical practice was also crucial. The vision is to develop simple-to-use tools for clinicians in hospitals, community practices and low-resource settings, to predict the risks of infection transmission and side-effects, optimise the choice of antimicrobials, and tailor dosage. Tools and models will also be critical for surveillance of AMR spread and transmission.

Need for large and diverse data sets

AI models are based on datasets that are used to train and subsequently validate them. Much of the data (e.g., demographics, clinical, microbiological, antimicrobial used) is collected as part of routine clinical practice and is contained in biobanks, medical records, and electronic health records. Research studies collect additional supplementary data (e.g., for pharmacokinetic or microbiological analysis). However, such health data is complex and unstructured: data from different sources often differs in form and can be difficult to combine. Moreover, certain populations - including many at higher risk of infection - are greatly underrepresented in these datasets. This can result in missing data and consequent bias in AI models.

Throughout the day, there were calls for large-scale (global), diverse and centralised datasets, collected in a systematic and standardized way. In addition, there is a need for improved processes for long-term maintenance and curation of the data, and continuity of personnel and expertise. Further depth and value will be brought by improved diagnostics, real-time monitoring capabilities through novel sensors and data (e.g., omic, epidemiological, usage, and absorption). Natural-language processing models that can mine unstructured text in electronic medical records will also be valuable.

Optimising models

AMR can be heterogeneous at the level of both the individual isolate and the population; geographic and ethnic heterogeneity can also limit the applicability of one model. AI models must therefore consider variabilities across populations and account for variability in healthcare settings (e.g., different patient pathways), countries, and external factors such as vaccination schemes. They must use appropriately large datasets, tackle bias and must be validated on unseen datasets. Models will also be improved by a deeper understanding of the biological mechanisms of AMR, including genotype-phenotype relationships and the impact of specific genetic mutations, and the complex interactions between the host, the antimicrobial and the pathogen.  

Translation into clinical practice

The translation of AI applications to the practical management of AMR will require behavioural change (e.g. in prescribing habits and effective use of available tools). Therefore, consultation and collaboration with end-users is critical as tools are developed and deployed.

The use of AI to influence health decisions and clinical management also needs consideration of regulatory requirements, ethics, and the acceptability by both practitioners and patients. Regulatory approval and acceptance may take a long time; in the meantime, there is a critical need to continue efforts to prevent the emergence of AMR organisms and manage existing ones, e.g. through AI-informed drug discovery.

Conclusions

The incorporation of AI tools is an endeavour that requires a global collaborative effort by scientists, clinicians, regulators, and the public, to improve data, regulation, adoption, and behavioural change. Imperial researchers are already leading the way in many of these efforts and are involved in several international consortia, including CAMO-Net, BenchmarkDR, AMR-Lung, and VITAL. Developing tools to improve patient outcomes, reduce misuse, and contribute to stewardship and surveillance will be critical in combating the huge healthcare burden of AMR and hospital-acquired infections.

You can find the full line of speakers here. Thanks to all the speakers and Chairs for their contributions.