Associations of serum sTREM-1 and sTREM-2 with mortality and neurological prognosis in patients resuscitated from cardiac arrest: a machine learning-based approach

Background: Patients resuscitated from cardiac arrest (CA) commonly have poor outcomes with a high mortality rate. We aimed to determine the predictive values of serum soluble triggering receptor expressed on myeloid cells 1 and 2 (sTREM-1 and sTREM-2) in patients after return of spontaneous circulation (ROSC) and to develop machine learning (ML) prediction models. Methods: We prospectively enrolled adult CA patients successfully resuscitated after cardiopulmonary resuscitation between November, 2021 to December, 2023. Serum sTREM-1, sTREM-2 and other biomarkers were measured on days 1, 3 and 5 after ROSC. The primary outcome was 28-day all-cause mortality. The secondary outcome was 3-month neurological prognosis. The performance of serum sTREM-1, sTREM-2, as well as the developed ML prediction models, to predict 28-day all-cause mortality and 3-month neurological prognosis were studied. Results: The study enrolled 120 patients, including 32 survivors and 88 non-survivors, with 30 healthy volunteers. Both sTREM-1 and sTREM-2 levels increased in patients after ROSC, with a larger increase in the non-survivors than survivors. Moreover, eleven features, including sTREM-1 and sTREM-2, were ultimately identified to build ML models. Among other ML models, the eXtreme Gradient Boosting (XGBoost) and Random Forest (RF) models showed strong performances for predicting 28-day all-cause mortality and 3-month neurological prognosis, respectively. Conclusion: Serum sTREM-1 performed better than sTREM-2 to predict mortality and neurological outcome after ROSC. Furthermore, the newly developed XGBoost and RF models incorporating sTREM-1 and/or sTREM-2 demonstrated superior predictive accuracy compared to conventional clinical scoring systems.