Developing artificial intelligence (AI) for clinical antimicrobial stewardship in an era of increasing antimicrobial resistance (AMR).
Lead Research Organisation:
University of Birmingham
Department Name: Institute of Microbiology and Infection
Abstract
This project has a specific focus in managing the single greatest threat to global health, the increasing burden from infections caused by bacteria that are resistant to antibiotics (antimicrobial resistance, AMR).
Doctors (humans) can't reliably know which antibiotic to administer in an emergency. In fact, from our earlier work they get it wrong about 20% of the time by prescribing an antibiotic that bacteria are resistant to for certain common types of infection. A serious bacterial infection will look the same whether the causative bacteria are resistant to certain antibiotics or not, and the first antibiotic must be selected on very limited information and be given the first 'golden' hour of sepsis management if the patient shows signs of an infection that has spread through the body. Understandably, this 'high stakes' uncertainty promotes the use of broad-spectrum antibiotics which should be held in reserve for known drug-resistant infections. Microbiological confirmation of which antibiotics are effective, when available, takes time (typically 2-3 days) which is too late for minimising un-necessary drug exposure and is often disregarded since the patient is 'getting better' on their broad-spectrum antibiotic. For many severe infections we simply never know if AMR was present, because biological samples are not taken or they were taken after the initiation of antibiotics that sterilise these samples. In the absence of knowing any alternatives, broad-spectrum antibiotics are often continued for fear of AMR being present. The emergency room decision and choice of first antibiotic seems to be the single-most important decision, not just for surviving sepsis but also for the antimicrobial stewardship needed to tackle the increasing problem of AMR.
Computer science has the potential to safely unlock successful antimicrobial stewardship for AMR at the first dose. Most AMR infections arise from 'gram-negative' bacteria that live in the gut, biliary and urinary systems, so in earlier work we used linked clinical and microbiological datasets from patients who needed emergency hospital admission for these pathogens in the blood and urine. The first step was to look at which antibiotics were given at emergency presentation, how often a patient was prescribed an antibiotic that their bacterial infection was resistant to (under-prescribing), and how often a broad-spectrum antibiotic was used when another, narrow-spectrum, antibiotic alternative would have been equally effective (over-prescribing). Using a patient's electronic health record (EHR), a computer system (or artificial intelligence, AI) trained in finding patterns in vast amounts of data, that was allowed to under-prescribe at the same rate as doctors (about 20% of the time), could also reduce the use of broad-spectrum antibiotics by about 40% by anticipation of which patients were unlikely to have an AMR infection. This powerful proof-of-concept work shows the huge potential for AI in making stepwise changes towards personalised medicine and antimicrobial stewardship at the first and most important dose (full paper doi:10.1093/jac/dkaa222).
Taking the next steps in AI for AMR. Many other ('gram-positive') bacteria that typically live on the skin and elsewhere are capable of serious infection and AMR. We can develop, test and model AI algorithms against these too, and broaden the potential for AI to help with more emergency hospital presentations. And even though many infections are not confirmed microbiologically, we can accurately infer information from the EHR and apply what an AI-supported prescribing pattern would look like. Linkage with clinical outcomes would also give a wider assessment of the impact AI might have on patient care and healthcare resources. Ultimately, we need a tool for front-line clinicians to safely prescribe but reduce any un-necessary initiation of broad-spectrum antibiotics.
Doctors (humans) can't reliably know which antibiotic to administer in an emergency. In fact, from our earlier work they get it wrong about 20% of the time by prescribing an antibiotic that bacteria are resistant to for certain common types of infection. A serious bacterial infection will look the same whether the causative bacteria are resistant to certain antibiotics or not, and the first antibiotic must be selected on very limited information and be given the first 'golden' hour of sepsis management if the patient shows signs of an infection that has spread through the body. Understandably, this 'high stakes' uncertainty promotes the use of broad-spectrum antibiotics which should be held in reserve for known drug-resistant infections. Microbiological confirmation of which antibiotics are effective, when available, takes time (typically 2-3 days) which is too late for minimising un-necessary drug exposure and is often disregarded since the patient is 'getting better' on their broad-spectrum antibiotic. For many severe infections we simply never know if AMR was present, because biological samples are not taken or they were taken after the initiation of antibiotics that sterilise these samples. In the absence of knowing any alternatives, broad-spectrum antibiotics are often continued for fear of AMR being present. The emergency room decision and choice of first antibiotic seems to be the single-most important decision, not just for surviving sepsis but also for the antimicrobial stewardship needed to tackle the increasing problem of AMR.
Computer science has the potential to safely unlock successful antimicrobial stewardship for AMR at the first dose. Most AMR infections arise from 'gram-negative' bacteria that live in the gut, biliary and urinary systems, so in earlier work we used linked clinical and microbiological datasets from patients who needed emergency hospital admission for these pathogens in the blood and urine. The first step was to look at which antibiotics were given at emergency presentation, how often a patient was prescribed an antibiotic that their bacterial infection was resistant to (under-prescribing), and how often a broad-spectrum antibiotic was used when another, narrow-spectrum, antibiotic alternative would have been equally effective (over-prescribing). Using a patient's electronic health record (EHR), a computer system (or artificial intelligence, AI) trained in finding patterns in vast amounts of data, that was allowed to under-prescribe at the same rate as doctors (about 20% of the time), could also reduce the use of broad-spectrum antibiotics by about 40% by anticipation of which patients were unlikely to have an AMR infection. This powerful proof-of-concept work shows the huge potential for AI in making stepwise changes towards personalised medicine and antimicrobial stewardship at the first and most important dose (full paper doi:10.1093/jac/dkaa222).
Taking the next steps in AI for AMR. Many other ('gram-positive') bacteria that typically live on the skin and elsewhere are capable of serious infection and AMR. We can develop, test and model AI algorithms against these too, and broaden the potential for AI to help with more emergency hospital presentations. And even though many infections are not confirmed microbiologically, we can accurately infer information from the EHR and apply what an AI-supported prescribing pattern would look like. Linkage with clinical outcomes would also give a wider assessment of the impact AI might have on patient care and healthcare resources. Ultimately, we need a tool for front-line clinicians to safely prescribe but reduce any un-necessary initiation of broad-spectrum antibiotics.
Technical Summary
The University of Birmingham works closely with University Hospitals Birmingham NHS Foundation Trust (UHB), the single-largest Acute NHS Trust in the UK that serves the healthcare needs of over 1.2m people in the second-largest city in the UK. PIONEER, the Health Data Research Hub for Acute Care, includes >1.2m patient episodes per year with >10yrs longitudinal data, 27.7m measurements/samples related to hospitalised infection and 573k prescriptions for sepsis.
We will use data collected from 2010-17 to train the algorithm, and data from 2018-22 to validate the algorithm's performance (data from 2020-21 will be analysed separately). Inclusion criteria will be acute admissions with confirmed bacterial infections. Training data will consist of the patient's electronic health record (EHR) at the time of patient admission, local resistance pressure (% resistance in samples collected in the last 90 days), and prescribing pressure (hospital-wide antibiotic consumption in the last 90 days). Outcome to be predicted will be the resistance profile (across 6 drugs for gram-negative pathogens and 3 for gram-positive) and the narrowest spectrum antibiotic that the infection is predicted to be sensitive to. We will compare the performance of several popular AI algorithms and a simple decision tree. Measures of success will be the % correct predictions of resistance across antibiotics being considered, % correct recommendation of antibiotic, and rates of over/under-prescribing. We will contrast the performance of AI and clinician recommendations across these measures, and where decisions differ we will review a subset with clinicians.
We will examine the performance of the AI across species, Gram status, infection type, age, ethnicity and socioeconomic strata of the patient. We will also examine the importance of variables from the EHR, and the relative importance of hospital-level variables (resistance and prescribing rates) versus individual patient-level information.
We will use data collected from 2010-17 to train the algorithm, and data from 2018-22 to validate the algorithm's performance (data from 2020-21 will be analysed separately). Inclusion criteria will be acute admissions with confirmed bacterial infections. Training data will consist of the patient's electronic health record (EHR) at the time of patient admission, local resistance pressure (% resistance in samples collected in the last 90 days), and prescribing pressure (hospital-wide antibiotic consumption in the last 90 days). Outcome to be predicted will be the resistance profile (across 6 drugs for gram-negative pathogens and 3 for gram-positive) and the narrowest spectrum antibiotic that the infection is predicted to be sensitive to. We will compare the performance of several popular AI algorithms and a simple decision tree. Measures of success will be the % correct predictions of resistance across antibiotics being considered, % correct recommendation of antibiotic, and rates of over/under-prescribing. We will contrast the performance of AI and clinician recommendations across these measures, and where decisions differ we will review a subset with clinicians.
We will examine the performance of the AI across species, Gram status, infection type, age, ethnicity and socioeconomic strata of the patient. We will also examine the importance of variables from the EHR, and the relative importance of hospital-level variables (resistance and prescribing rates) versus individual patient-level information.