Truace
Subscribe

Model flags which patients with positive blood cultures may die within 30 days, study finds

Study title: Machine learning prediction of 30-day mortality in adults with a positive blood culture

Researchers trained nine machine learning algorithms on records of adults in the MIMIC-IV database who had at least one positive blood culture, to predict who would die within 30 days. Of 5,731 episodes, 742 patients (12.9%) died within that period.

Read the study abstract

Introduction A positive blood culture is a clinically heterogeneous finding that may represent true bloodstream infection, transient bacteremia or contamination-diagnostic uncertainty that complicates early clinical decision-making. We aimed to develop and evaluate machine learning models to predict 30-day all-cause mortality in adults with a positive blood culture. Methods Retrospective cohort study using MIMIC-IV. Adult hospitalizations with at least one positive blood culture were included (first positive cul…

Truvace record2 min read1 sourceEvidence: Moderate · 63/100 — Observational study · n = 5,731 · 1 independent sourcePeer-reviewedTRV-2026-2418 · cite or check it
Machine learning prediction of 30-day mortality in adults with a positive blood culture
"Blood Culture Bottles" by geekmoose is licensed under CC BY 2.0. To view a copy of this license, visit https://creativecommons.org/licenses/by/2.0/.

In brief

The random forest model performed best after threshold optimization, with sensitivity 0.788, specificity 0.707 and balanced accuracy 0.748. Across all nine models, test-set ROC AUC ranged from 0.759 to 0.801.

A positive blood culture can mean true bloodstream infection, transient bacteremia or contamination, which complicates early decisions. The authors say predicting risk at the moment of culture positivity, rather than waiting for a sepsis diagnosis, could support early risk stratification.

Main points

  1. The study was a retrospective cohort analysis of MIMIC-IV, using the first positive blood culture per adult admission.
  2. Of 5,731 episodes, 742 patients (12.9%) died within 30 days.
  3. Nine supervised algorithms were trained with repeated 10-fold cross-validation and evaluated in an independent test cohort.
  4. Test-set ROC AUC ranged from 0.759 to 0.801 across models.
  5. The random forest had a Brier score of 0.0999 and a calibration slope of 1.03.
  6. The main predictors were ICU admission at onset, mechanical ventilation, bilirubin, urea and the Charlson comorbidity index.
  7. Potential contaminants were kept in the data to reflect real-world uncertainty.

The gain

In a retrospective study of 5,731 hospital episodes, a random forest model predicted 30-day death after a positive blood culture with 0.788 sensitivity and 0.707 specificity using routine data.

The full story

The team included adult hospitalizations with at least one positive blood culture, counting only the first positive culture per admission. Potential contaminants were retained so the data would reflect the diagnostic uncertainty clinicians face.

Predictors were demographic, clinical, laboratory, organ support and microbiological variables. Nine supervised algorithms were trained with repeated 10-fold cross-validation, tested in an independent cohort and calibrated, with thresholds set using the Youden index.

Of 5,731 episodes, 742 patients (12.9%) died within 30 days. Test-set ROC AUC ranged from 0.759 to 0.801, and the random forest showed the best overall balance of sensitivity (0.788) and specificity (0.707).

The random forest was well calibrated, with a Brier score of 0.0999 and a calibration slope of 1.03. ICU admission at onset, mechanical ventilation, bilirubin, urea and the Charlson comorbidity index were the leading predictors.

Sources

  1. Peer-reviewedEnfermedades infecciosas y microbiologia clinica (English ed.)2026-11-01

The debate