AI in Antibiotic Stewardship

Antibiotic stewardship programs are coordinated efforts to make sure patients get the right antibiotic, dose and duration, and no antibiotic when none is needed. In the United States, Joint Commission standards and federal Medicare conditions of participation require hospitals to run them. Common strategies include prospective audit and feedback, where pharmacists and infectious disease physicians review active prescriptions and suggest changes. Another is preauthorization, where certain drugs need approval before use.

The hardest moment is empiric therapy: choosing a drug before culture results arrive. The traditional aid is the antibiogram, a hospital-wide table showing what share of each organism was susceptible to each drug. But an antibiogram is an average and says nothing about the person in front of you. A patient with two prior resistant urine cultures and recent ciprofloxacin exposure is very different from one with no history.

Machine learning models personalize that estimate. They use prior cultures, recent antibiotics, hospitalizations, nursing home residence, procedures and other conditions. In a 2020 study in Science Translational Medicine, Sanjat Kanjilal and colleagues in Boston trained models on outpatient urinary tract infections. Tested retrospectively, their algorithm could have reduced the use of second-line antibiotics while keeping the rate of ineffective therapy similar to or lower than clinicians' choices.

A second line of work speeds up the lab itself. MALDI-TOF mass spectrometry already identifies bacterial species quickly. Researchers, including a Swiss team that built the DRIAMS dataset, have trained models to predict resistance from the same spectra, which could narrow therapy sooner.

A common misconception is that these tools cut antibiotic use by themselves. A resistance model cannot tell whether a patient needs antibiotics at all. It cannot tell a true infection from asymptomatic bacteriuria, for example. And under pressure to treat possible sepsis, clinicians may add broad coverage anyway. The tools work best as inputs to a stewardship team's judgment, not as replacements for it.