Machine-learning algorithms identifies sTREM1 has a key biomarker for outcome prediction in critically ill
Charles de Roquetaillade1,2,3, Pierre-Louis Blot4,5, Fabrice Uhel6,7
1Université Paris Cité, Inserm U942 MASCOT, Paris, F-75006, France. charles.de-roquetaillade@aphp.fr.
Critical Care (London, England)
|May 28, 2026
Summary
Soluble triggering receptor expressed on myeloid cells-1 (sTREM-1) is a strong predictor of mortality and kidney outcomes in critically ill patients. This biomarker performs comparably to machine learning models and surpasses traditional severity scores.
Area of Science:
- Critical care medicine
- Biomarker discovery
- Machine learning in healthcare
Background:
- Traditional prognostic tools for critically ill patients have limitations in capturing disease complexity.
- Existing methods like severity scores and single biomarkers offer incomplete prognostic insights.
Purpose of the Study:
- To evaluate the prognostic capability of multiple biomarkers, alone and with clinical data, using machine learning.
- To predict mortality and kidney-related outcomes in intensive care unit (ICU) patients.
Main Methods:
- Post-hoc analysis of the FROG-ICU cohort (n=2,061) and external validation in the MARS cohort.
- Assessed 15 plasma biomarkers and employed Random Forest (RF) and LASSO regression models.
- Determined variable importance using RF's mean decrease in accuracy.
Main Results:
- Machine learning models achieved an Area Under the Curve (AUC) of 0.74 for mortality prediction, outperforming severity scores (AUC 0.64).
- Soluble triggering receptor expressed on myeloid cells-1 (sTREM-1) emerged as the strongest predictor (AUC 0.72), performing comparably to ML models.
- sTREM-1 showed similar high performance for predicting major adverse kidney events (MAKE).
Conclusions:
- sTREM-1 is a reliable biomarker for predicting mortality and kidney outcomes in critically ill patients.
- Its prognostic performance rivals advanced machine learning models and exceeds traditional scoring systems.
