Prediction of Antimicrobial Resistance in People Living With Cystic Fibrosis Using Machine Learning

ABSTRACT

Antimicrobial resistance (AMR) is a growing challenge in people living with cystic fibrosis (pwCF), who often experience chronic lung infection and repeated antibiotic exposure. Because routine antibiotic susceptibility testing often takes several days, treatment is frequently started before current resistance profiles are available. We assessed whether routinely collected electronic healthcare records (EHR) could predict antibiotic resistance in sputum cultures from adults with CF. In this retrospective single‐center study, 12,618 sputum cultures from 209 pwCF between 2012 and 2022 were linked with 63,823 days of intravenous antibiotic exposure, spirometry, demographic characteristics, microbiology results, and historical resistance data. Different models were trained and evaluated with patient‐level splitting and cross‐validation to predict resistance to ciprofloxacin, ceftazidime, meropenem, piperacillin/tazobactam, and tobramycin. Extreme gradient boosting showed the most consistent performance with AUCs of 0.75–0.80. Model discrimination was broadly similar in cultures with and without Pseudomonas aeruginosa, except for ceftazidime and meropenem. Shapley Additive Explanations (SHAP) suggested that longer term resistance history was more informative than recent results. These findings support the feasibility of using EHR‐derived data to estimate AMR before culture results are available, but external validation, broader antibiotic exposure data, and assessment of temporal dataset shift are needed before clinical use.