Machine learning clustering of multidimensional clinical features identifies four prognostically distinct sepsis subtypes

Sepsis is highly heterogeneous, and traditional “one-size-fits-all” treatment approaches are insufficient for optimal risk stratification. Using the MIMIC-III database, we employed unsupervised machine learning (k-means clustering) based on eight key clinical features (age, SOFA score, lactate, urine output, systolic blood pressure, heart rate, creatinine, and BUN) to identify clinically distinct subtypes among 4,559 sepsis patients, with clustering validated via principal component and silhouette analyses. Four distinct subtypes were identified: Young Low-risk (30.4%; 30-day mortality 7.7%), Elderly Stable (49.4%; 18.0%), High-risk (9.7%; 25.9%), and Critical (10.6%; 54.3%). Significant differences were observed among subtypes in age distribution, organ dysfunction severity, and prognosis (log-rank P < 0.001). Cluster assignment added statistically significant prognostic information beyond SOFA score and age (AUC increment 0.0126, DeLong P = 3.34 × 10⁻⁵; category NRI 0.0330, 95% CI 0.0056–0.0543). Cluster analysis based on multidimensional clinical characteristics can effectively identify sepsis patient subtypes with distinct prognoses, providing a pragmatic framework for risk stratification in sepsis; the incremental prognostic value beyond conventional severity scores warrants validation in prospective cohorts.