The discovery of antibiotics nearly a century ago revolutionized the treatment of infectious diseases around the world. But today, 700,000 people die each year from infections that have become resistant to antibiotics—and microbial mortality is expected to jump dramatically to 10 million by 2050.

In the United States, the CDC reports that 28% of antibiotic prescriptions by doctors and emergency departments are unnecessary. This growing public health threat is due in part to clinicians overprescribing, patient pressure and uncertain diagnoses.

“Your doctor is just making an educated guess—and we tend to guess way bigger than we need to, bringing out a giant shotgun when just a small scalpel would work,” says Jonathan Chen, MD, PhD, an assistant professor of medicine (biomedical informatics) at Stanford whose research is attempting to combat the growing threat of antimicrobial resistance (AMR.)

“Everyone paid attention to COVID, a tsunami that overran the world,” said Chen, a former Stanford Health Policy VA Medical Informatics Fellow. “But antibiotic resistance is like rising sea levels: you barely notice it but eventually it’s going to kill more and more people.”

Chen and his Stanford colleague Mary K. Goldstein, MD, a professor of health policy and a former bioinformatics researcher at the U.S. Department of Veterans Affairs, are using data science to help doctors make better informed decisions and health-care facilities to adopt a precision stewardship approach to combatting antimicrobial resistance.

“We’re using machine learning to make use of extensive electronic health record (EHR) data that otherwise is not ordinarily available, in a summarized format to be able to guide prescription use of antibiotics,” Goldstein said. Electronic patient charts already provide myriad data that can be fed into predictive models, such as the history of past infections and antibiotic susceptibility, past prescriptions, medical history and lab results.

Personalized AI

Chen, equipped with an NIH grant, is building personalized “antibiograms” with machine-learning tools that can predict antibiotic susceptibility for individuals based on patterns learned from large collections of prior examples. Artificial intelligence (AI) is already embedded in our daily lives, writes Chen in this paper published in the Journal of Antimicrobial Chemotherapy. Instead of fearing it, he says, why not harness data from electronic health records to predict the best and safest use of antibiotics?

“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,” Chen writes. “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.”

In one study, the researchers developed machine learning models that predict antibiotic susceptibility patterns using EHR data of 8,342 infectious from Stanford emergency departments and 15,806 urinary tract infections from Massachusetts General Hospital and Brigham & Women’s Hospital in Boston. The study was published in Communications Medicine, a Nature Portfolio journal. The researchers created AI “antibiograms” using EHR data to predict antibiotic susceptibility, measure the performance of antibiotic selections informed by the antibiograms compared with the selections made by clinicians, and then evaluate the trade-off in performance when fewer broad-spectrum antibiotics were selected across a specific population of patients. They took a cohort of patients from Stanford emergency departments between 2009 and 2019 and then replicated the process on a similar cohort of patients from Massachusetts General Hospital and Brigham & Women’s Hospital in Boston between 2007 and 2017.

In both study cohorts, had antibiotics been prescribed based on machine-learning based predictions, just as many patients would have received appropriate treatment while using far less broad-spectrum antibiotics. For example, the personalized antibiogram approach would have allowed 69% of the patients receiving vancomycin + piperacillin/tazobactam to have instead received piperacillin/tazobactam alone while achieving the same coverage of patient infections. Similarly, personalized antibiogram guidance could have resulted in 40% of patients receiving piperacillin/tazobactam to have instead received a much more targeted antibiotic (cefazolin).

On the other hand, Chen explains, the initial human-driven prescriptions were not broad enough, covering the patient’s actual infection only 84% to 88% of the time. The use of personalized antibiogram guidance could have increased that coverage rate to 86% to 90% of the time.

“This precision approach to prescribing brings a win-win result that doesn’t just offer a tradeoff between safety and stewardship, but a means to improve both simultaneously,” Chen said.

Social Disparities in Resistance

Other predictors of antimicrobial resistant are social determinants of health, namely one’s Zip code, education, economic stability, the surrounding environment and access to health care. Using the Area Deprivation Index, which ranks neighborhoods by socioeconomic disadvantages—with a ranking of 100 being the most disadvantaged—Chen found that areas with high ADI are more likely to higher rates of organisms resistant to antibiotics.

In a new study published in the journal of Clinical Infectious Diseases, Chen and his colleagues collected patient bacterial culture results from 2015 to 2020 from EHRs of two large health-care systems within the Dallas-Fort Worth Texas metropolitan area. They used geocoded cultures and linked them to socioeconomic index values. They found significant clusters of AMR organisms in areas with high levels of deprivation, particularly for AmpC—enzymes resistant to penicillin and other antibiotics— and MRSA, a potentially fatal bacteria now resistant to an entire class of antibiotics.

“Interventions that improve socioeconomic factors such as poverty, unemployment, decreased access to healthcare, crowding, and sanitation in these areas of high prevalence may reduce the spread of AMR,” the research team wrote.