Machine-learning algorithms identify sTREM1 as a key biomarker for outcome prediction in critically ill Authors: Charles de Roquetaillade, Pierre-Louis Blot, Fabrice Uhel, Louis Boutin, Jérôme Cartailler, Tom Van Der Poll, Etienne Gayat, Alexandre Mebazaa, Benjamin Chousterman Published: 2026 ABSTRACT AbstractIntroductionPrognostic assessment in critically ill patients traditionally relies on severity scores or single biomarkers, each with limited ability to capture the biological heterogeneity of critical illness.ObjectiveTo compare the prognostic performance of multiple biomarkers, individually and in combination with clinical variables, using machine learning approaches for the prediction of mortality and kidney-related outcomes.Materials and methodsWe performed a post-hoc analysis of the FROG-ICU cohort, a prospective observational study of patients admitted to ICUs. The study included critically ill patients who required invasive mechanical ventilation or a vasoactive agent for more than 24 h. The primary outcome was day-90 mortality, secondary outcome was major adverse kidney event (MAKE) in ICU. A total of 15 plasma biomarkers were evaluated using multiparametric approach. ML models involved Random Forest (RF) and LASSO regression. Mean decrease in accuracy was used to determine variable importance in RF model. External validation was performed in the MARS cohort which involved ICU patients admitted for sepsis and septic shock.ResultsAmong 2,061 patients in the FROG-ICU day-90 mortality was 30.1%. Machine learning models achieved AUCs of 0.74, outperforming severity scores (AUC 0.64, p < 0.001). Variable importance analysis consistently identified sTREM-1 as the strongest predictor. When evaluated alone, sTREM-1 demonstrated high prognostic performance (AUC 0.72), comparable to ML models. These findings were confirmed in the MARS cohort. Similar results were observed for MAKE prediction.ConclusionsTREM-1 is a robust biomarker associated with mortality and kidney-related outcomes in critically ill patients. Its predictive performance were comparable to multiparametric machine learning models and superior to severity scores.Supplementary InformationThe online version contains supplementary material available at https://doi.org/10.1186/s13054-026-06092-9.Keywords: Machine learning; ICU; Sepsis, mortality; MAKE; Treatment response prediction; Conserved immune dysregulation Introduction Prognostic assessment in ICU patients traditionally relies on severity scores or the evaluation of a single biomarker [1, 2]. While severity scores have been widely used, they are based on basic biological assessment and clinical parameters, therefore, they may not capture the underlying pathophysiological mechanisms driving disease severity and progression. These mechanisms, often referred to as the “host response”, are widely recognized as major determinants of organ dysfunction in critically ill. Severity scores are also very sensitive to premorbid condition and age [3] and, for some of them, were developed decades ago and may no longer reflect evolving demographics, therapeutic practices, and prognostic of ICU populations [2, 4]. In recent years, a large number of biomarkers have been investigated for their prognostic value in specific critical care settings, including sepsis, trauma, and postoperative care. These biomarkers typically reflect specific pathophysiological pathways such as systemic inflammation, endothelial activation, coagulation abnormalities, or organ-specific dysfunction (e.g., renal or hepatic injury). Despite their individual relevance, most biomarkers have been studied in isolation and within narrowly defined clinical contexts (sepsis, acute respiratory distress syndrome, trauma) [5–7]. As a result, few studies have explored their comparative prognostic performance or evaluated their potential additive value when assessed together across heterogeneous ICU populations [8]. Recent insights from multimodal monitoring studies, have challenged the conventional approach which consists in stratifying patients based on primary admission diagnosis [9, 10]. These studies support a shift from phenotype-based classifications toward biologically informed endotypes [10, 11]. In this evolving context, there is a growing need to redefine critical illness based on biology rather than clinical syndromes [12]. In this context, machine learning approaches offer the opportunity to integrate multidimensional biological and clinical data, enabling the identification of the most informative predictors of outcome without prespecified assumptions. Such approaches may help move beyond syndrome-based classifications toward biologically informed prognostic frameworks. Materials and methods Study design We used the FROG-ICU (French and European Outcome Registry in ICUs) cohort as an exploratory cohort [13]. For validation purpose, we used the MARS-ICU cohort, which enrolled ICU patients admitted for sepsis and septic shock [14]. The FROG-ICU study was a multicentric, international prospective, observational study conducted in 24 centers from five countries in Europe from June 2015 to May 2016. Endpoint definition The primary objective was to compare the performance of multiple candidate biomarkers when taken either independently or collectively and to examine their association with day 90 mortality. Secondary outcomes included the association between the tested biomarkers and major adverse kidney event (MAKE) in ICU [15]. MAKE was defined by the presence of any of the following: death in ICU; need for renal replacement therapy during ICU stay or non-recovery defined as increase in creatinine at discharge > 1,25 x baseline creatinine. Baseline creatinine was defined as previously described [16]: if the estimated glomerular filtration rate (eGFR) on admission exceeded 75 mL/min/1.73 m², the admission creatinine value was used. If eGFR was < 75 mL/min/1.73 m², baseline creatinine was estimated by back-calculating from the MDRD equation, assuming an eGFR of 75 mL/min/1.73 m². Biomarker measurements Blood was collected within 24 h after inclusion. Samples were subsequently processed and stored at -80°c before transfer to the central laboratory for analysis. Measurements were performed on Cobas® e801 module in EDTA plasma via Elecsys® ElectroChemiLuminescence (ECL) technology, A total of 15 biomarkers were measured, a detail of the biomarkers and their function is presented in Supplementary Table 1. Biomarkers were selected based on their involvement in key pathophysiological pathways that drive organ failure and mortality in critically ill patients. Namely, interleukin-6 (IL-6), soluble Triggering Receptor Expressed on Myeloid cells-1 (sTREM-1), C-reactive protein (CRP), solubleCD14 (presepsin), calprotectin, Growth Differentiation Factor-15 (GDF-15), and soluble Urokinase (suPAR) reflect distinct components of the innate inflammatory response. We chose soluble Fms-like tyrosine kinase-1 (sFLT-1), heparin-binding protein (HBP), Angiopoietin-2 (Ang-2), endothelial cell-specific molecule-1 (ESM-1), and tyrosine kinase receptor (Tie-1) that capture endothelial activation and vascular permeability, a central determinant of shock and multiorgan dysfunction. Protein induced by vitamin K absence or antagonist-II (PIVKA-II) reflects sepsis-associated coagulopathy; and insulin-like growth factor binding protein-7 (IGFBP-7) is a validated marker of early kidney stress and injury. Our aim was to cover the major mechanistic domains implicated in sepsis and critical illness, immune activation, endothelial dysfunction, coagulation abnormalities, and organ-specific damage, to identify which biological axis contributes most to prognostic stratification. Statistical analyses Quantitative variables are presented as medians with interquartile ranges (IQRs), and qualitative variables as counts with percentages. Biomarker concentrations were log-transformed and analyzed as continuous variables. Categorical variables were compared using the Chi-squared test (with Yates’ continuity correction for binary variables). For continuous variables, we used either the student’s t test for normalized variables or the Mann–Whitney U test for non-normal variables. Statistical significance was defined as p < 0.05. Missing data were handled via multiple imputation using the predictive mean matching method. Patients with > 3 missing biomarkers or missing day-90 mortality data were excluded. Colinearity was assessed between variables (Supplementary Fig. 1). Due to high collinearity between Presep-A and Presep-B (R² > 0.95), only Presep-B was retained for analysis. The FROG-ICU dataset was randomly split into a training set (70%) and a testing set (30%). Model development, including variable selection and hyperparameter tuning, was performed exclusively on the training set. Model performance (AUC) was evaluated on the testing set. External validation was subsequently performed in the MARS cohort. Three models were applied: Least Absolute Shrinkage and Selection Operator (LASSO) logistic regression, Random Forest (RF), and logistic regression (LR). For the RF model, the number of trees was 500, variable importance was assessed using mean decrease accuracy, and Gini as a sensitivity analysis. Clinical variables included in the models corresponded to those used to compute the SAPS II score, which was the only severity score consistently available across both cohorts. To ensure comparability across variables with different scales, all continuous predictors were standardized prior to analysis. Standardization was performed using z-score normalization, where each value was centered by subtracting the mean and scaled by dividing by the standard deviation. This transformation results in variables with a mean of 0 and a standard deviation of 1, facilitating model convergence and interpretation of coefficients. All the tested biomarkers were visually inspected to ensure that no aberrant outlier was present. To identify the most important variables associated with the outcome, we used LASSO logistic regression with the glmnet package in R. This method selects the most relevant predictors by reducing the influence of less important ones, and automatically setting some coefficients to zero. We applied cross-validation to choose the best penalty value (lambda) that balances model simplicity and prediction accuracy. The variables with non-zero coefficients at this optimal lambda were considered important. These selected variables were then ranked based on the absolute value of their coefficients, which reflects their relative contribution to the model. RF is known for high predictive accuracy, provided variable importance metrics via mean decrease in accuracy and Gini index. LR was used to evaluate associations between individual biomarkers and outcome. Models’ performances were assessed using receiver operating characteristic (ROC) curves and corresponding areas under the curve (AUCs). Comparison between AUCs’ was performed using the DeLong test. To further illustrate the association between sTREM-1 and day-90 mortality, survival analyses were performed after dichotomizing patients according to the optimal sTREM-1 threshold. This threshold was derived from the training subset of the FROG-ICU cohort using receiver operating characteristic (ROC) analysis and the Youden index. Kaplan–Meier survival curves were generated and compared using the log-rank test. Hazard ratios (HRs) with 95% confidence intervals were estimated using Cox proportional hazards models. Survival time was calculated from ICU admission to death or censoring at day 90. Analyses were conducted using R version 4.4.2 with the “glmnet” and the “randomForest” packages. Results After exclusion of patients with missing data, a total of 2061 patients from the FROG-ICU cohort and 251 patients from the MARS cohort could be included in our analyses (see Flow-Chart in Supplementary material). Day 90 mortality was 30.1% and 37.8% respectively (Table 1). Table 1Baseline characteristics and outcomes between survivors and non survivors in the FROG-ICU cohort and the MARS cohort stratified according to day 90 mortalityData are presented as median [IQR] or n (%). Comparisons between survivors and non-survivors were performed using appropriate statistical tests, with significant p-values highlighted. ICU-free days at day 28: Number of days within the first 28 days after admission during which the patient is alive and not in the ICU. Patients who die before day 28 are assigned 0 ICU-free days. MAKE (Major Adverse Kidney Events): Composite endpoint including death, need for renal replacement therapy, or persistent renal dysfunction at follow-up (as defined in the original cohort criteria). ICU-mortality: Death occurring during the ICU stay. Day-90 mortality: All-cause mortality at 90 days Baseline characteristics and outcomes between survivors and non survivors in the FROG-ICU cohort and the MARS cohort stratified according to day 90 mortality Data are presented as median [IQR] or n (%). Comparisons between survivors and non-survivors were performed using appropriate statistical tests, with significant p-values highlighted. ICU-free days at day 28: Number of days within the first 28 days after admission during which the patient is alive and not in the ICU. Patients who die before day 28 are assigned 0 ICU-free days. MAKE (Major Adverse Kidney Events): Composite endpoint including death, need for renal replacement therapy, or persistent renal dysfunction at follow-up (as defined in the original cohort criteria). ICU-mortality: Death occurring during the ICU stay. Day-90 mortality: All-cause mortality at 90 days Primary outcome The distribution, values, correlation and detailed description of all 15 tested biomarkers are presented in Supplementary Tables 2 and Supplementary Fig. 1. Using these biomarkers in combination with clinical variables, the LASSO regression model achieved an AUC of 0.74 [0.70–0.76] for the prediction of 90-day mortality (Fig. 1A). A comparable performance was observed with the RF algorithm, which yielded an AUC of 0.74 [0.70–0.78] (p = 0.63 for comparison). Both machine learning models significantly outperformed logistic regression based on clinical severity scores alone, which demonstrated an AUC of 0.64 [0.60–0.68] (p < 0.001 for comparison). When each biomarker was evaluated individually, sTREM1 showed the highest prognostic performance, with an AUC of 0.72 [0.68–0.77] for 90-day mortality prediction. Its performance was not significantly different from that of the LASSO and Random Forest models (p = 0.28 and 0.35, respectively), but was superior to SAPS2 score (p < 0.01). Validation in the MARS cohort confirmed the prognostic performance of sTREM1 for predicting 90-day mortality, with an AUC of 0.71 [0.64–0.77]. This was comparable to the performance of the SAPS II score, which showed an AUC of 0.69 [0.62–0.76] (Fig. 1B). Receiver-operating characteristic (ROC) curves for predicting day-90 mortality (A) and MAKE in ICU (B) based on a multiparametric approach (LASSO, Random Forest) or sTREM-1 alone and SAPS II In the Random Forest model, variable importance ranking based on mean decrease in accuracy identified sTREM1 as the most predictive variable, ahead of the clinical score (Fig. 2). Other top-ranking biomarkers included IGFBP7 and GDF15. Coefficient analysis from the LASSO model confirmed similar findings, with sTREM1 being the most strongly associated with 90-day mortality (Supplementary Table 3). Kaplan–Meier analysis confirmed a significantly higher day-90 mortality in patients with elevated sTREM-1 levels (log-rank p < 0.001) (Fig. 3). Mean decrease in accuracy plot from the Random Forest model illustrating the relative importance of variables (top to bottom) Kaplan–Meier curves for day-90 mortality according to sTREM-1 levels. Patients were categorized using the optimal cutoff derived from the training FROG-ICU cohort (1360pg/mL). Survival was compared using the log-rank test Secondary outcome We further assessed the predictive performance of selected biomarkers for secondary outcomes. MAKE during ICU stay occurred in 39.7% of patients in the FROG-ICU cohort and in 44.6% of patients in the MARS cohort (Table 1). In the FROG-ICU cohort, machine learning models demonstrated high discriminative ability for predicting MAKE. The LASSO regression model achieved an AUC of 0.80 [0.78–0.82], and the Random Forest model achieved an AUC of 0.81 [0.78–0.85] (p = 0.91 for comparison), indicating similar performance. Both models significantly outperformed SAPSII yielding an AUC of 0.66 [0.60–0.69] (p < 0.001 for comparisons). Variable importance analysis based on mean decrease in accuracy in the Random Forest model highlighted sTREM1 and Presepsin as the most predictive features (Supplementary Fig. 3). Similarly, coefficient analysis from the LASSO model identified sTREM1 as the strongest predictor, surpassing both the clinical score and other biomarkers (Supplementary Table 3). In the MARS validation cohort, similar findings were observed. sTREM1 demonstrated an AUC of 0.85 [0.80–0.89] for the prediction of MAKE in ICU, outperforming the SAPS II score, which had an AUC of 0.75 [0.68–0.80] (Fig. 4). Receiver-operating characteristic (ROC) curves for predicting MAKE in ICU in the FROG-ICU cohort (A) and in the MARS cohort (B) The optimal threshold of sTREM1 for predicting 90-day mortality was 1,360 pg/mL (sensitivity: 0.66; specificity: 0.69). For MAKE in ICU, the optimal threshold was 1,420pg/mL (sensitivity: 0.77; specificity of 0.68). sTREM1 association with outcome and subgroup analyses We examined the distribution of sTREM1 across patient subgroups and its association with clinical outcomes. At ICU admission, sTREM1 levels correlated moderately with SAPS II scores (R² = 0.36, Spearman correlation p < 0.01) but were log-linearly associated with 90-day mortality (p < 0.001) (Supplementary Fig. 2). Subgroup analyses showed a consistent association between sTREM1 concentrations (as a continuous variable) and 90-day mortality across sex, age groups, illness severity, presence of sepsis, and ARDS (Fig. 5). Similar associations were observed between sTREM1 and MAKE during ICU stay. Forest plot showing odds ratios (OR) for day-90 mortality and MAKE in ICU in subgroup analyses based on continuous sTREM-1 levels (OR reported per 1 log-transformed increase of sTREM-1) Discussion Taken together, our results identify sTREM-1 as a robust prognostic biomarker associated with mortality and kidney-related outcomes in critically ill patients. Importantly, sTREM-1 alone showed predictive performance comparable to multiparametric ML models, suggesting that a single biologically informative biomarker may capture a substantial proportion of prognostic information. Since its first description in 2000 [17] sTREM1 has been associated with outcome prediction, primarily in the context of sepsis [18, 19]. Here, we demonstrate its prognostic value across a range of admission diagnoses. These findings align with recent studies challenging the concept of disease-specific host-response pathways and instead supporting shared mechanisms across disease endotypes [7, 9, 10, 20]. TREM-1 is a transmembrane receptor expressed on innate immune cells (monocytes, granulocytes), as well as endothelial cells and platelets. Activation of the TREM-1 pathway synergizes with TLR signaling to amplify inflammation [21]. In critical illness, this amplification may contribute to dysregulated host responses, increased organ damage, and higher mortality [10]. However, our findings should be interpreted as demonstrating association rather than causation, and our study does not establish that modulation of the TREM-1 pathway would improve outcomes. Elevated TREM-1 signaling can be reliably assessed by measuring sTREM-1 [22]. Indeed, several authors have demonstrated that most of sTREM1 originated from proteolytic cleavage of mature membrane-anchored TREM-1 rather than alternative splicing of mRNA. Interest in the TREM-1 pathway has grown in recent years, notably because its inhibition has been proposed as a therapeutic strategy to dampen hyperinflammation in ICU patients. Nangibotide, a 12–amino acid peptide derived from TREM-like transcript 1 (TLT-1), can act as a decoy receptor and modulate TREM-1 signaling. Preclinical studies have demonstrated its anti-inflammatory effects, mostly in sepsis models [23–25]. In the recent ASTONISH trial, 402 septic shock patients were randomized to receive nangibotide or placebo [26]. Although no benefit was observed in SOFA score reduction, a prespecified exploratory analysis suggested a clinically meaningful benefit in patients with high baseline sTREM1 concentrations (> 532pg/mL). These findings suggest that sTREM-1 may help identify patients at higher risk, but whether targeting the TREM-1 pathway improves outcomes remains to be determined. Our study has several limitations. First, biomarker availability in the validation cohort precluded application of ML models to the MARS cohort; however, logistic regression yielded comparable performance between cohorts, strengthening our conclusions. Second, patient characteristics differed between cohorts: patients in MARS had sepsis exclusively, were younger, and had fewer comorbidities than those in FROG-ICU. Differences in age may partly explain the variation in sTREM-1 levels observed between cohorts, as prior studies have reported age-dependent variations in sTREM-1 concentrations [27]. In addition, differences in assay techniques may also have contributed to variations in absolute biomarker levels. In the FROG cohort, sTREM-1 was measured using an Elecsys® electrochemiluminescence assay, whereas a standard ELISA was used in the MARS cohort. Third, the set of clinical variables was limited to those included in SAPS II, and additional clinical parameters or biomarkers may have improved model performance. In particular, procalcitonin (PCT), a widely used marker of systemic inflammation, and neutrophil gelatinase-associated lipocalin (NGAL), a biomarker of kidney injury, were not available in both cohorts and therefore could not be evaluated. Finally, the FROG-ICU cohort was conducted more than a decade ago, and changes in ICU management over time may limit the generalizability of our findings. Conclusion Using a multiparametric machine learning approach, we identified sTREM-1 as a potential key biomarker for risk stratification in unselected ICU patients. Higher admission sTREM1 levels were consistently associated with mortality and kidney-related outcomes. Further studies are needed to assess whether targeting TREM-1 signaling in patients with elevated admission sTREM1 could improve outcomes. Supplementary Material Supplementary Material 1. Acknowledgements The authors used a large language model (ChatGPT-5) for language editing. All authors subsequently reviewed and revised the text and take full responsibility for the content of the manuscript. Author contributions CdR led the analysis and manuscript preparation. FU, CdR, and LB collected the data. AM and BGC supervised the study. Funding None. Data availability Data will be made available upon reasonable request formulated to the corresponding author. Declarations Ethical approval and consent to participate The original FROG-ICU study was conducted in France and Belgium in accordance with Good Clinical Practice (Declaration of Helsinki 2002) and Ethical Committee approvals (Comité de Protection des Personnes—Ile de France IV, IRB n°00003835 and Commission d’éthique biomédicale hospitalo-facultaire de l’hôpital de Louvain, IRB n°. Consent for publication All authors have read the paper and gave their consent for publication. Competing interests The assays were performed as part of the IMOCRIT research program, funded by Roche. BGC served as a member of a Roche Diagnostics advisory board and received no other fees. CdR received fees from Roche Diagnostics. Roche Diagnostics had no role in study design, data collection, data analysis, manuscript preparation, or the decision to submit the manuscript. The other authors declare no competing interests in relation to the present study. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. Change history 7/9/2026 The title contained minor errors in the original publication, this article has been updated. References Associated Data Supplementary Materials Supplementary Material 1. Data Availability Statement Data will be made available upon reasonable request formulated to the corresponding author.