SHBC1090

A.LI¹, T. M. NG², A. CHOW² 

MOH Holdings Pte Ltd (MOHH)¹, Tan Tock Seng Hospital² 

Early identification of bacteremia is a critical first step in guidance of prompt, appropriate and prudent antimicrobial therapy. Artificial Intelligence (AI) techniques can complement conventional regression models in identifying early critical predictors of bacteremia.

We included a cohort of 39,029 patients who were admitted to Tan Tock Seng Hospital and had blood cultures done, 2016-2017. Of which, n=4,350 (11.1%) patients were confirmed to have bacteremia with positive blood cultures.

We selected 38 features from demographical, clinical, vital signs and laboratory data available within the first 12 hours of admission, as inputs into models predicting for bacteremia. Following AI models were explored: eXtreme Gradient Boosting (XGBoost), Gradient Boosting (GBoost), Random Forest (RanFor), MultiLayer Perceptron (MLP) neural networks. Logistic regression (LogReg) model was applied as a baseline for comparison. Repeated 5 fold cross validation and training were performed on 70% of the cohort. Model performance was tested and compared on 30% of the cohort. SHapley Additive exPlanations (SHAP) was applied on the best performing AI model for interpretation of early predictors of bacteremia.

XGBoost was found to be the best performing model with Area Under Receiver Operator Curve of 0.797 (LogReg: 0.754, GBoost: 0.791, RanFor: 0.789, MLP NN: 0.773) and Area Under Precision Recall Curve of 0.412 (LogReg: 0.362). SHAP identified the following top predictive features and their associated value range: Procalcitonin (>1.03ug/L), %Neutrophils (>83.7%), temperature (>37.5?) and CRP (>42.9mg/L).

High performing AI models can be applied on routinely available clinical data for  bacteremia prediction in the first 12 hours of admission.