Predicting Community-Onset Candidemia in an Academic Medical Center Using Machine Learning

• Published in: Infection control and hospital epidemiology 2020-10, Vol.41 (S1), p.s355-s355
• Main Authors: AlZunitan, Mohammed, Marra, Alexandre, Edmond, Michael, Street, Nick, Diekema, Daniel, Salinas, Jorge
• Format: Article
• Language: English
• Subjects: Antibiotics; Antifungal agents; Blood; Cardiovascular diseases; Corticoids; Dementia disorders; Diabetes; Disease; Disease control; Drug abuse; Electronic health records; Fungal infections; Hypertension; Machine learning; Patients; Public health; Sexually transmitted diseases; STD
• ISSN: 0899-823X; 1559-6834
• DOI: 10.1017/ice.2020.974

Background: Candidemia is a leading cause of bloodstream infections (BSIs), and community-onset candidemia is being recognized as a public health problem. In the era of electronic health records (EHRs), we can use machine learning to detect patterns in patient data that may predict infections. Objective: We aimed to predict community-onset candidemia in patients admitted to the University of Iowa Hospital & Clinics (UIHC) using machine-learning algorithms. Methods: We retrospectively reviewed data for patients admitted to UIHC during 2015–2018. All adult inpatients who had a requested blood culture were included. Candidemia was defined as a blood culture positive for Candida within 48 hours after admission. Variables of interest were extracted from the EHR: age, sex, body mass index, and month of admission. We also included comorbidities upon admission defined by the International Classification of Diseases, 10th Revision, Clinical Modification (ICD-10-CM): cardiovascular diseases, neurological disorders, chronic pulmonary disease, dementia, rheumatoid disease, peptic ulcer disease, liver disease, diabetes mellitus, hypothyroidism, renal failure, coagulopathy, obesity, weight loss, fluid and electrolyte disorders, anemia, alcohol abuse, drug abuse, psychiatric diseases, malignancy, and HIV/AIDS. We calculated Charlson and Elixhauser scores based on ICD-10-CM codes. We also included prehospitalization conditions (90 days before admission): Candida- positive cultures from sites other than blood, antibiotics/antifungals, hemodialysis, central lines, corticosteroids, surgeries, and intensive care unit (ICU) admissions. Mode and median imputation were used for missing information. Random forests with resampled training sets were used for prediction, and results were evaluated using 10-fold cross validation. Results : In total, 30,528 adult admissions were extracted; 73 admissions had an episode of candidemia (