BSAC Vanguard Series: Artificial intelligence and antibiotic stewardship

Abstract

Antimicrobial stewardship is a key facet in preventing antimicrobial resistance but can be difficult to put into practice. Frontline providers are faced with the unknowns of pending culture data plus the urgency of appropriate antibiotic choice to prevent sepsis-related mortality; this often leads to broad-spectrum antibiotic prescribing. Currently available resources lack a customized approach to individual patients. Artificial intelligence (AI) focused on antimicrobial stewardship may create a unique opportunity to provide individualized, real-time recommendations to providers on appropriate, but narrower spectrum, antibiotic options. We envision that, with further advances in AI, personalized clinical decision support tools to optimize antibiotic prescribing could be available within the next decade.

Introduction

Antibiotics are one of the most important advances in modern medicine, increasing the average human lifespan by over 20 years. However, increasing bacterial resistance is a natural and inevitable consequence of antibiotic use, requiring essential strategies to sustain these gains in health. The driving concept of antibiotic stewardship is using the right antibiotic, at the right dose, at the right time. This collides with data suggesting that up to half of all antibiotic prescriptions are either unnecessary or inappropriate.

Why is appropriate use of antibiotics so hard?

The choice of antibiotics when first encountering a patient, known as ‘empiric’ antibiotic prescribing, is essentially educated guesswork. Antibiotic appropriateness is determined by the infecting pathogen and its susceptibility to various antibiotics. Unfortunately, the key tests providing these data, microbial cultures, often take days to provide results. This is far too long to delay treatment that could minimize the chance of the patient dying from sepsis. If we as doctors don’t know what we’re treating, but we know we need something that works immediately or else the patient might die, then we often give an antibiotic that treats everything possible. This mentality leads to physicians picking a ‘shotgun’ broad-spectrum antibiotic, when a narrow-spectrum precision ‘scalpel’ would have been more effective and less damaging. Moreover, many non-infectious diseases can look similar to infection, making decisions even harder as to whether any antibiotic is needed at all.

National guidelines on specific syndromes (e.g. pneumonia, skin/soft tissue infection) provide general suggestions on when to treat infection and what antibiotics to use, but they can take years to develop and can only provide overly general one-size-fits-all guidance that providers are often not even aware of. Local hospital antibiograms document resistance patterns within a specific healthcare facility, but are still not specific to particular patients. We need better tools at the point of care to guide precision antibiotic prescribing, personalized and rapidly updated using new streams of data.

Artificial intelligence and machine learning for antibiotic stewardship?

Existing tools and standards to support antibiotic prescribing provide important guidance but are too broad for personalized recommendations that must balance the immediate risks of undertreatment against the nebulous risks of overtreatment. This actionable, arbitrary, and ascertainable process where an important decision (empiric antibiotic prescribing) depends on our variable human ability to predict a verifiable result (diagnostic culture susceptibilities) is ideally suited for innovative machine learning methods.

Artificial intelligence (AI) is often depicted as science fiction of a distant future, yet modern algorithms are already shaping our lives every day. Do internet advertisements seem uncannily specific to your interests? This is the power of predictive analytics: using machine learning predictive models on large-scale data to generate individualized predictions and suggestions. Imagine if we were to use this same power, except for choosing the right antibiotic, individualized to a single patient. With similar technology, we can use the vast amounts of data provided by electronic medical records to create predictive models to optimize the accuracy and consistency of the current ‘educated guesswork’ of empiric antibiotic prescribing.

Electronic patient charts provide ample information that can be utilized by models, ranging from the history of past infections and antibiotic susceptibility data, to patient-specific presenting symptoms, past medical history, laboratory results, and imaging. We currently use this data in large scientific studies to better determine which variables independently predict the need for specific types of antibiotics in general. Using the right tools, AI models can rapidly sift through similar data, and with statistical training, they can be taught to predict the answers to specific, patient-oriented questions. Does this particular patient have an infection or not? If so, is it safer to use the ‘shotgun’ or the ‘scalpel’ approach? If we give our AI the right variables and the right training, then models can estimate the probabilities of infection with antimicrobial-resistant bacteria for specific patients quantitatively in a manner that humans can only intuit qualitatively. The output can be real-time, giving providers the guidance needed at a critical time when it is needed the most. Just as we would expect from competent clinicians, AI models should not be stagnant. Continuously learning algorithms should adapt to incoming streams of new patient and epidemiologic data. With additional training, predictions can become better and more relevant, allowing for a more dynamic guidance tool than static guidelines or yearly antibiograms.

There have already been breakthroughs in this area. For instance, at Stanford we are pitting a machine learning model’s suggestions for empiric antibiotics against the choices already made by our own providers. We compared the predictions made for antibiotic selection at the beginning of an infection with the resistance profiles from culture information obtained at the end. Such models can choose a correct antibiotic based on resistance profile as well, or better, than the clinician’s choice while choosing narrower-spectrum antibiotics. This indicates the potential to improve both safety and stewardship, simultaneously. Similar work is emerging in the form of locally trained and optimized tools in multiple sites.

Making an impact at population scales will require advances in exchanging data across diverse health systems, the key fuel to power modern AI systems. While substantial inertia tends to lock such data into local health system silos, the shock of the COVID pandemic finally motivated many to pool their data, resources, and effort to face a larger threat. Laws on improving interoperability between health systems as well as international standards such as Fast Healthcare Interoperability Resources (FHIR) are advancing efforts in this area. Hopefully we will collectively recognize the even greater, but slowly growing, problem of antimicrobial resistance and take action. Within the next decade, we envision providers routinely using AI models as their own personal guidance tools, ones which provide recommendations that are not just effective for treating immediate infections, but also in combating the otherwise relentless development of antimicrobial resistance.

Transparency declarations

This article first appeared as one of a series of blog posts celebrating the fiftieth anniversary of the founding of the British Society for Antimicrobial Chemotherapy. Dr Chen heads the HealthRex research group in the Stanford Center for Biomedical Informatics Research, seeking to combine human and artificial intelligence approaches to medicine that will deliver better care than either can alone. Dr Chang has none to declare.