Machine learning based prediction of antibiotic sensitivity in patients with critical illness

Rising antibiotic resistance inflicts a heavy burden on healthcare, both clinically and economically. Owing to the time required to obtain culture and sensitivity test results, quite often the clinicians rely on their experience and static clinical guidelines to prescribe antibiotics. Such empirical treatment often fails to account for patient-specific attributes and changes in the antibiotic resistance patterns with time and location. The aim of this study was to analyze the patient and hospital specific features regarding their prognostic relevance to treat bacterial infections of patients in the intensive care units (ICUs). We performed a single-center retrospective cohort analysis across 25526 positive cultures recorded in MIMIC-III critical care database. We retrieved a number of clinically relevant relationships from association analysis between patient factors and bacterial strains. For instance, higher elapsed time from patient admission to sample collection for culture showed strong association with blood stream infection caused by Enterococcus faecium, Pseudomonas aeruginosa, and Staphylococcus, indicating that these infections are possibly hospital acquired. To predict antibiotic sensitivity at the level of individual patients we developed an ensemble of machine learning algorithms. The model provided superior prediction accuracy (about 87%) and area under the ROC curve (around 0.91 on an average) for the four most common sample types as compared to a number of off-the-shelf techniques. We demonstrate the predictive power of commonly recorded patient attributes in personalised prediction of antibiotic efficacy.