Development of Machine Learning Models to Predict Candidemia in Hospitalized Adult Patients

We aimed to develop a generalizable machine learning model that leverages electronic medical record (EMR) data to predict candidemia using information available on or before the date of blood culture collection.

This study included adult patients admitted to Shin Koga Hospital (April 2014-March 2022) who underwent blood cultures at least 14 days after admission. We prepared two datasets: a "Complete case dataset" (13 variables, no missing values) and a "Imputed full-variable dataset" (36 variables, single imputation). We evaluated the discriminative performance of four machine learning models, including extreme gradient boosting, random forest, least absolute shrinkage and selection operator (LASSO) logistic regression, and logistic regression using stratified 5-fold cross-validation by comparing their area under the receiver operating characteristic curve (AUROC). The model with the highest mean AUROC was selected, assessed on the entire dataset, and internally validated using 1,000 bootstrap replicates.

A total of 919 patients were included, of whom 47 (5.1%) had candidemia. The LASSO logistic regression demonstrated the best performance and was selected as the final model. The AUROC (95% CI) was 0.880 (0.836-0.924) for the Complete case dataset and 0.913 (0.878-0.948) for the Imputed full-variable dataset. The optimism-corrected AUROCs were 0.851 and 0.876, respectively.

We developed two versions of the LASSO logistic regression model using only information automatically extracted from EMR, considering real-world data availability. These models have the potential to serve as practical tools to support early diagnosis and therapeutic intervention for candidemia.