Machine-Learning algorithms identifies sTREM1 has a key biomarker for outcome prediction in a mixed-ICU population

Introduction : Prognostic risk assessment in ICU patients traditionally relies on severity scores or the evaluation of a single biomarker. Although multiple biomarkers have been proposed for this purpose, comparative analysis of their prognostic performance is limited. Machine Learning (ML) methods offer the ability to integrate numerous biomarkers together with clinical variables simultaneously and may help identify those most effective for prognosis assessment. Objective : This study sought to investigate the predictive performance of a panel of inflammatory biomarkers and clinical variables both individually or in combination using ML algorithms for mortality prediction in a mixed-ICU population. Secondary objectives included the prediction of organ-specific outcomes, specifically major adverse kidney events (MAKE). Materials and Methods : This study was a post-hoc analysis of the French and EuRopean Outcome ReGistry in Intensive Care Units (FROG-ICU) study, a prospective observational study of patients admitted to ICUs. The study included patients consecutively admitted to the ICU who required invasive mechanical ventilation or a vasoactive agent for more than 24 h. The primary outcome was day-90 mortality, secondary outcome was MAKE in ICU. A total of 14 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. Results : A total of 2,061 patients from the FROG-ICU cohort were analyzed. Among them, 621 (33.2%) died by day-90. Using a combination of 14 biomarkers together with severity scores, ML algorithms demonstrated similar areas under the curve (AUC) for predicting day-90 mortality (AUC 0.73 for LASSO and 0.74 for RF), outperforming severity scores with an AUC of 0.64 for SAPS II (p<0.01 for comparison). Variable importance analysis, based on mean decrease in accuracy, identified sTREM-1 as a key biomarker for risk stratification. When assessed independently, sTREM-1 showed high performance for day-90 mortality prediction (AUC 0.72) comparable to the ML algorithms. Similar results were observed for kidney-related outcomes, with AUCs of 0.78 for sTREM-1, 0.80 for LASSO, and 0.84 for Random Forest. External validation in the MARS cohort confirmed our findings. Conclusion : Through a multiparametric approach using ML, we identified sTREM-1 as a potential key biomarker for risk stratification in ICU patients, predicting both mortality and kidney-related outcomes, consistently across admission diagnoses. sTREM-1 provides superior prognostic information as compared to established severity scores. Further research is required to determine whether targeting this pathway could benefit ICU patients.