Identification of Clinical Subtypes of Surgical Sepsis in Critically Ill Patients Based on K-Means Clustering and Development of Parsimonious Classifier Model for High-Risk Subtype.

Background: Identifying the phenotypes of patients with surgical sepsis is crucial for improving their management. This study aims to identify the clinical subtypes of critically ill surgical sepsis patients and develop parsimonious classifier model for rapid subtype identification. Methods: A retrospective cohort study was conducted on 4,027 adult surgical sepsis patients admitted to the Surgical Intensive Care Unit at the First Affiliated Hospital of Sun Yat-sen University from 2018 to 2023, with external validation using the MIMIC-IV (the Medical Information Mart for Intensive Care) database. We extracted the worst values of clinical variables within 24 hours before and after diagnosis and performed K-means clustering. The elbow method and the silhouette coefficient was used to determine the optimal number of clusters. Survival analysis was used to evaluate the prognostic differences among different subtypes. A multivariate logistic regression model was constructed to identify the subtypes, and the model was evaluated by the area under the receiver operating characteristic curve (AUROC), decision curve analysis (DCA), calibration curve and SHAP (Shapley Additive Explanations). Results: Three subtypes were finally determined. Subtype-1 (13%) exhibited severe organ dysfunction and the highest 1-year mortality (73.3%), subtype-2 (59%) had mild organ dysfunction and the lowest 1-year mortality (26.4%), while subtype-3 (28%) was characterized by older age and prevalent comorbidities. External validation confirmed similar subtype distributions and clinical profiles in the MIMIC-IV cohort. A four-variable logistic regression model (including prothrombin time (PT), dosage of norepinephrine, lactate and pH) was developed to predict the high-risk subtype-1. The model demonstrated satisfactory discriminative ability in the derivation cohort (AUC: 0.970) and robust performance in external validation (AUC: 0.889). SHAP analysis highlighted PT, lactate, and dosage of norepinephrine as core predictors of subtype-1. Conclusion: This study identified three clinical subtypes of critically ill patients with surgical sepsis based on early clinical variables. A multivariate logistic regression model was developed to identify the high-risk subtype-1, with core impact factors including PT, dosage of norepinephrine and lactate.