Infection and Drug Resistance (Feb 2021)
Developing Machine-Learning Prediction Algorithm for Bacteremia in Admitted Patients
Abstract
Ebrahim Mahmoud,1 Mohammed Al Dhoayan,2,3 Mohammad Bosaeed,1,4,5 Sameera Al Johani,5,6 Yaseen M Arabi5,7 1Department of Infectious Disease, Department of Medicine, King Abdulaziz Medical City, Riyadh, Saudi Arabia; 2Department of Health Informatics, CPHHI, King Saud Bin Abdulaziz University for Health Sciences, Riyadh, Saudi Arabia; 3Data and Business Intelligence Management Department, ISID, King Abdulaziz Medical City, Riyadh, Saudi Arabia; 4King Abdullah International Medical Research Center (KAIMRC), Riyadh, Saudi Arabia; 5College of Medicine, King Saud Bin Abdulaziz University For Health Sciences, Riyadh, Saudi Arabia; 6Department of Pathology & Laboratory Medicine, King Abdulaziz Medical City, Riyadh, Saudi Arabia; 7Department of Intensive Care, King Abdulaziz Medical City, Riyadh, Saudi ArabiaCorrespondence: Ebrahim MahmoudDivision of Infectious Diseases, Department of Medicine, King Abdulaziz Medical City, Riyadh, Riyadh, Saudi ArabiaTel +966 500081418Email [email protected]: Bloodstream infection among hospitalized patients is associated with serious adverse outcomes. Blood culture is routinely ordered in patients with suspected infections, although 90% of blood cultures do not show any growth of organisms. The evidence regarding the prediction of bacteremia is scarce.Patients And Methods: A retrospective review of blood cultures requested for a cohort of admitted patients between 2017 and 2019 was undertaken. Several machine-learning models were used to identify the best prediction model. Additionally, univariate and multivariable logistic regression was used to determine the predictive factors for bacteremia.Results: A total of 36,405 blood cultures of 7157 patients were done. There were 2413 (6.62%) positive blood cultures. The best prediction was by using NN with the high specificity of 88% but low sensitivity. There was a statistical difference in the following factors: longer admission days before the blood culture, presence of a central line, and higher lactic acid—more than 2 mmol/L.Conclusion: Despite the low positive rate of blood culture, machine learning could predict positive blood culture with high specificity but minimum sensitivity. Yet, the SIRS score, qSOFA score, and other known factors were not good prognostic factors. Further improvement and training would possibly enhance machine-learning performance.Keywords: bacteremia, blood culture prediction, machine learning, predictive medicine
