Prediction of bacterial and fungal bloodstream infections using machine learning in patients undergoing chemotherapy

Abstract

Purpose This study aimed to develop a machine learning (ML) model to predict bloodstream infection (BSI) in chemotherapy patients.

Patients and methods We included all cancer patients undergoing chemotherapy at a tertiary cancer hospital from 2017 to 2022. Data were collected per chemotherapy cycle, including chemotherapy drugs, indications, cycle number, cancer type, body mass index, age, gender, complete blood count, creatinine levels, and microbial cultures. BSI was assessed within 21 days after chemotherapy. The ML algorithms tested included logistic regression, ridge regression, k-nearest neighbors, Naive Bayes, Perceptron, neural networks, decision trees, boosting methods, Random Forests, and Support Vector Machines. The SHapley Additive exPlanations (SHAP) method was used to measure feature importance.

Results Among 107,757 cycles from 19,225 patients, 91.7 % had solid tumors, primarily breast (36.8 %) and gastrointestinal (19.4 %) cancers. The first cycle accounted for 23.7 % of cycles, and palliative chemotherapy made up 52.9 %. Alkylating agent was the most common drug class used (55.5 %). BSI occurred in 1.33 % of cycles, with 34 % of these cases occurring in neutropenic patients. Of the bacteremia cases, 11.8 % were polymicrobial, and 69.3 % involved gram-negative bacteria. The best model was a neural network with one hidden layer (5 neurons), achieving 70.7 % sensitivity, 93.49 % specificity, 93.19 % accuracy, and an area under a receiver operating characteristic curve of 91.93 %. Key predictors included the first cycle, antimetabolite use, palliative chemotherapy, monocytopenia, and hematological malignancies.

Conclusion ML effectively predicts bacteremia in chemotherapy patients, including non-neutropenic cases, and could be used in clinical practice to guide treatment and infection workup.