Explainable machine learning for discontinuation of therapeutic antibiotics in intensive care patients.

INTRODUCTION: The threshold for initiating lifesaving antibiotics for intensive care patients is low while determining when to stop remains challenging: prolonged administration increases antibiotic resistance and occurrence of side effects, while premature discontinuation may lead to resurgence of infection. Explainable machine learning may predict for which ICU patients antibiotics will be reinitiated. METHODS: We retrospectively collected data on all adult ICU patients treated with antibiotics in two tertiary academic hospitals in the Netherlands. We included monitor data, laboratory data, selective decontamination strategy, medication, and culture results as predictors. We trained logistic regression, lightGBM and AutoPrognosis models to predict the primary outcome of restarting antibiotics within 72 h. RESULTS: We included data on 2,486 patients receiving 3,645 therapeutic antibiotic courses between October 2015 and January 2022. The primary outcome occurred 708 (19.4 %) times, of which the same subgroup was restarted in 468 (66.10 %) cases and the exact same combination in 278 (39.27 %). Cultures were collected in 80.4 % and positive in 43.3 % of cases. 90-day mortality was higher in the reinitiation group (39.8 % vs. 25.0 %). The best performing model was logistic regression (AUROC 0.675). The most important predictors were the use of penicillins and the treatment duration of the last started antibiotic. CONCLUSION: Although prediction of reinitiation remained challenging, the most important and consistent predictor for reinitiation of antibiotics was a shorter duration of the last started antibiotic. The same antibiotic was restarted frequently, and 90-day mortality was higher in the reinitiation group, illustrating potential for data-driven decision support on when to discontinue antibiotics.