目的 探索中药治疗脓毒症的“药-效”关联，识别与降低病死率、改善急性生理与慢性健康评分II（acute physiology and chronic health evaluation Ⅱ，APACHE Ⅱ）等关键临床结局相关的核心中药，为中药辅助治疗脓毒症提供循证参考。方法 系统检索截至2024年9月的9个中英文数据库，纳入中药复方治疗脓毒症的临床研究，构建脓毒症中药复方数据库。采用十折交叉验证比较9种机器学习模型的预测性能，并运用沙普利加性解释（Shapley additive explanations，SHAP）方法解析各临床结局的最优模型，识别关键药物及其贡献方向。结果 多层感知器在预测中医证候评分、有效率、炎症与免疫指标、生化指标、器官功能障碍评分及病死率性能最佳；逻辑回归在血气分析和胃肠功能与肠黏膜屏上表现最优；支持向量机在血常规与APACHE II评分上预测效能最佳。SHAP分析显示，地黄、枳实、黄芪、附子、黄芩、厚朴等中药与病死率、APACHE II评分、炎症指标和胃肠功能等结局中具有较高正向贡献，位于“药-效”网络的核心节点。结论 通过多模型比较与可解释性分析，构建了以临床结局为导向的中药“药-效”关联网络，揭示了核心中药在脓毒症治疗中的潜在关键作用，为中医药精准辨证用药及中药新药研发提供数据支撑和循证依据。

Objective To explore the “medicine-efficacy” associations of traditional Chinese medicine (TCM) in the treatment of sepsis, identify core herbsrelated to reducing mortality and improving key clinical outcomes such as the acute physiology and chronic health evaluation II (APACHE II) score, and provide evidence-based references for TCM-assisted treatment of sepsis. Methods A systematic search of Chinese and English databases was conducted up to September 2024 to collect clinical studies on TCM compound prescriptions for sepsis. A comprehensive sepsis TCM formula database was established. Nine machine learning algorithms were compared using ten-fold cross-validation, and the optimal model for each clinical outcome was interpreted through the Shapley additive explanations (SHAP) method to identify key herbs and their contribution directions. Results The multilayer perceptron showed the best performance in predicting TCM syndrome scores, overall effectiveness, inflammatory and immune indicators, biochemical parameters, organ dysfunction scores, and mortality; logistic regression performed best for blood gas analysis along with gastrointestinal function and intestinal mucosal barrier outcomes; and the support vector machine achieved optimal predictive performance for routine blood tests and APACHE II scores. SHAP analysis revealed that Dihuang (Rehmanniae Radix), Zhishi (Aurantii Fructus Immaturus), Huangqi (Astragali Radix), Fuzi (Aconiti Lateralis Radix Praeparata), Huangqin (Scutellariae Radix), and Houpo (Magnoliae Officinalis Cortex) had positive contributions across outcomes such as mortality, APACHE II score, inflammatory markers, and gastrointestinal function, forming the core nodes of the “medicine-efficacy” network. Conclusion This study established a clinically oriented “medicine-efficacy” association network through multi-model comparison and explainable machine learning analysis. The findings highlight the potential key roles of several core herbs in improving major clinical outcomes of sepsis, providing data support and evidence-based basis for precise syndrome differentiation and medication in TCM, as well as for the research and development of new TCMs.

TP18;R285.64

国家杰出青年科学基金项目(82225049);国家自然科学基金面上项目(82574870);四川省中医药管理局中医药科研专项(2024zd023);四川省中医药管理局中医药科研专项(25ZDAZX008);四川大学华西医院学科卓越发展1·3·5工程项目(ZYGD23004)