Development and validation of dynamic clinical subphenotypes using vital sign trajectories in stroke patients in the intensive care unit: A retrospective study based on the MIMIC-IV and eICU-CRD databases

Background Stroke constitutes a major contributor to morbidity and mortality in ICU settings. However, prognostic variability and insufficient dynamic assessment frameworks hinder timely therapeutic interventions. Accurate early outcome prediction is essential for enhancing clinical management and patient survival rates. Methods We performed a retrospective investigation utilizing MIMIC-IV and eICU-CRD databases. Multi-trajectory modeling was employed to characterize dynamic clinical phenotypes in stroke populations. Demographic and clinical parameters were extracted and comparative analyses conducted across phenotypic groups. Multiple machine learning algorithms were developed to determine optimal predictive performance and identify critical variables. Additional subgroup evaluations and survival assessments were undertaken. Results Four distinct dynamic clinical phenotypes were delineated. Substantial heterogeneity in clinical parameters was evident across these subgroups, each demonstrating distinctive pathophysiological profiles regarding inflammatory markers, metabolic parameters, and hemodynamic characteristics. The SGBT algorithm achieved superior discriminative capacity for in-hospital mortality prediction. Nine pivotal predictive features emerged: leukocyte count, blood glucose, body temperature, BUN, SpO₂, heart rate, body weight, respiratory frequency, and hematocrit. The SGBT model attained robust AUC values across all phenotypes. The trajectory-derived subphenotypes exhibited marked variations in physiological parameter patterns and clinical attributes, demonstrating consistent prognostic stratification capabilities for mortality assessment. Conclusion Nine critical predictive indicators for stroke outcomes were established. This investigation provides practical frameworks for risk stratification and therapeutic decision-making in stroke management.