Platelet trajectory subphenotypes in intensive care patients with sepsis-induced coagulopathy

We aimed to identify early platelet trajectory-based subphenotypes in sepsis-induced coagulopathy (SIC), assess their prognostic value, explore whether glucocorticoid-related mortality associations differed across trajectory groups given within 24 h of ICU admission, and develop an early prediction model for trajectory membership. We retrospectively analyzed 6151 SIC patients from MIMIC-IV and externally validated the findings in 1305 patients from eICU. Platelet trajectories during the first 96 h were identified using group-based trajectory modeling among patients with sufficient platelet monitoring. The primary outcome was 28-day all-cause mortality after ICU admission. Multivariable Cox regression was used to evaluate associations between trajectory groups and outcome. An exploratory analysis assessed possible heterogeneity in the association between early systemic glucocorticoid exposure and mortality. A four-class prediction model was developed from first-day variables using XGBoost and externally validated. Four platelet trajectory groups were identified in MIMIC-IV, accounting for 23.3%, 35.3%, 32.4%, and 9.0% of patients. In the fully adjusted model, Groups 1, 2, and 3 had higher 28-day all-cause mortality risk than Group 4, with Group 1 showing the highest risk. The adjusted HRs for Groups 1, 2, and 3 versus Group 4 were 2.53, 1.73, and 1.39 in MIMIC-IV, and 5.78, 3.37, and 2.49 in eICU, respectively. The exploratory glucocorticoid analysis suggested possible heterogeneity in glucocorticoid-related mortality estimates across trajectory groups. The 24 h prediction model showed good multiclass performance in both internal and external validation, with macro-AUCs of 0.873 in MIMIC-IV and 0.884 in eICU. Among SIC patients with sufficient platelet monitoring, early platelet trajectories identified prognostically distinct subphenotypes. The observed differences in glucocorticoid-related mortality associations should be interpreted cautiously as hypothesis-generating. The 24 h prediction model may support early risk stratification.