What question did this study set out to answer?

The aim is to create and validate a machine learning model that predicts positive blood cultures using only patients' vital signs.

April 27, 2026Open Access

Development and validation of a machine learning model for predicting positive blood cultures using vital signs in ICU patients

Key Points

The aim is to create and validate a machine learning model that predicts positive blood cultures using only patients' vital signs.
Utilized clinical data from ICU patients at two university hospitals, with 597 blood cultures for model development and 366 for validation.
Constructed a Balanced Random Forest model using six explanatory variables from vital signs measured over three days.
Evaluated model performance using area under the receiver operating characteristic curve for both internal and external validation.
Achieved an area under the curve of 0.700 ± 0.072 for internal validation.
Achieved an area under the curve of 0.679 ± 0.010 for external validation.
Identified key factors affecting predictions, including temperature range, blood pressure changes, and heart rate variations.

Abstract

Bloodstream infections require timely and appropriate diagnosis and treatment, as inadequate management is associated with higher mortality. Previous predictive models for bloodstream infection have generally incorporated laboratory test results in addition to vital signs, although laboratory results are not always immediately available in clinical practice. This study aimed to develop and validate a machine learning model to predict positive blood cultures using only vital signs in febrile intensive care unit (ICU) patients (maximum BT ≥ 38 °C). Clinical data from ICU patients at two university hospitals were included, with 597 blood cultures used for model development and 366 for external validation. Six explanatory variables derived from body temperature, heart rate, and mean arterial pressure measured over three days were used to construct a Balanced Random Forest model. The area under the receiver operating characteristic curve was 0.700 ± 0.072 for internal validation and 0.679 ± 0.010 for external validation. SHapley Additive exPlanations value analysis indicated that a larger temperature range contributed to positive predictions, while no increase in temperature, increases in blood pressure, and a decrease in heart rate contributed to negative predictions. The model demonstrates moderate predictive performance and shows comparable results across different datasets. It can continuously monitor the need for blood cultures, thereby serving as an adjunctive tool to support clinical decision-making.

AIに質問

Bookmark

View Full Paper