Receive a weekly summary and discussion of the top papers of the week by leading researchers in the field.

In Journal of biomedical informatics ; h5-index 55.0

Unnecessary antibiotic regimens in the intensive care unit (ICU) are associated with adverse patient outcomes and antimicrobial resistance. Bacterial infections (BI) are both common and deadly in ICUs, and as a result, patients with a suspected BI are routinely started on broad-spectrum antibiotics prior to having confirmatory microbiologic culture results or when an occult BI is suspected, a practice known as empiric antibiotic therapy (EAT). However, EAT guidelines lack consensus and existing methods to quantify patient-level BI risk rely largely on clinical judgement and inaccurate biomarkers or expensive diagnostic tests. As a consequence, patients with low risk of BI often are continued on EAT, exposing them to unnecessary side effects. Augmenting current intuition-based practices with data-driven predictions of BI risk could help inform clinical decisions to shorten the duration of unnecessary EAT and improve patient outcomes. We propose a novel framework to identify ICU patients with low risk of BI as candidates for earlier EAT discontinuation. For this study, patients suspected of having a community-acquired BI were identified in the Medical Information Mart for Intensive Care III (MIMIC-III) dataset and categorized based on microbiologic culture results and EAT duration. Using structured longitudinal data collected up to 24, 48, and 72 hours after starting EAT, our best models identified patients at low risk of BI with AUROCs up to 0.8 and negative predictive values >93%. Overall, these results demonstrate the feasibility of forecasting BI risk in a critical care setting using patient features found in the electronic health record and call for more extensive research in this promising, yet relatively understudied, area.

Eickelberg Garrett, Nelson Sanchez-Pinto L, Luo Yuan

2020-Aug-16

Antibiotic Stewardship, Critical Care, Electronic Health Records, MIMIC, Machine Learning, Prediction Models