深層時間グラフクラスタリングモデルを用いた臓器相互作用軌道に基づく敗血症のサブフェノタイピング：後ろ向きコホート研究

Sepsis is a heterogeneous syndrome with varying degrees of multi-organ dysfunction. Identifying dynamic inter-organ interactions is critical for accurate sepsis subphenotyping and targeted therapy, yet remains unexplored. In this study, we aimed to quantify the dynamic trajectories of organ interactions to define sepsis phenotypes, supporting personalized treatment and clinical decision-making.

We proposed a novel deep temporal graph clustering model to identify sepsis phenotypes by quantifying dynamic multi-organ interactions within 48 h post-diagnosis. The model was trained and validated on the Medical Information Mart for Intensive Care III (MIMIC-III) dataset (admissions from 2001 to 2012) and externally validated on the eICU Collaborative Research (eICU) dataset (admissions from 2014 to 2015). Its effectiveness was benchmarked against state-of-the-art clustering algorithms. Patient characteristics, multi-organ system states coupling patterns, and prognostic outcomes were compared across the identified phenotypes. Extreme gradient boosting (XGBoost) was used for early phenotype classification at 4 h post-diagnosis. To enhance clinical applicability, a user-friendly web interface was developed. Propensity score matching and weighted logistic regression were employed to evaluate the effects of the fluid management strategies on in-hospital mortality of patients with various phenotypes.

A total of 10,181 and 6208 unique sepsis patients were employed as the cohorts for the model development and external validation, respectively. Three distinct phenotypes were identified and labeled as Phenotype A, B, and C, exhibiting significant differences in baseline characteristics, organ system states coupling patterns, and outcomes (P-value < 0.05). Phenotype A had the lowest mortality (5.59%) and accounted for the largest proportion of patients (46.34%). In contrast, Phenotype C represented the highest mortality (38.27%) and comprised the smallest proportion (22.78%). Phenotype A was characterized by sustained synchronous improvement across organ system states. Phenotype B showed persistent decoupling of organ system states. Phenotype C exhibited a rapid transition from early asynchrony to synchronization. The model demonstrated robust clustering performance in external validation. The simplified classifier showed high predictive performance, achieving an area under the receiver operating characteristic curve (AUROC) of 0.84 (95% CI [0.83, 0.86]) for phenotype prediction at 4 h post-diagnosis. The beneficial fluid management strategies varied across different phenotypes, highlighting the need for targeted fluid intervals.

This study characterizes sepsis phenotypes using organ interaction trajectories and identifies three heterogeneous patterns of disease progression. These patterns offer new insights into the underlying pathophysiological mechanisms of sepsis, which can support the design of clinical trials on disease progression and guide the optimal allocation of intensive care resources.

This study was funded by the National Nature Science Foundation of China (No. 32371372) and the National Key Research and Development Program of China (No. 2022YFC2009503).