应用机器学习（ML）建立脓毒性休克患者住院28 d死亡预测模型，并验证其有效性。

对2019年1月至2024年6月苏州大学附属苏州九院收治的脓毒性休克患者的临床资料进行回顾性分析。根据患者入院28 d的预后分为死亡组（104例）和生存组（161例）。使用最小绝对收缩选择算子（LASSO）算法筛选脓毒性休克患者住院28 d死亡的关键变量。总体数据以3∶1比例随机分配到训练集和测试集。通过比较8种不同的机器学习算法，选择最优算法来重建脓毒性休克患者住院28 d死亡的预测模型。使用受试者工作特征（ROC）曲线下面积（AUC）和临床决策曲线（DCA）评估模型的性能。Shapley加性解释（SHAP）算法用于模型解释。

共纳入265例脓毒性休克患者。通过临床初步筛选，模型最初包含47个变量，经过LASSO模型进一步筛选后，确定9个变量用于后续模型开发。在构建的8个模型中，径向基函数支持向量机（RSVM）模型表现出最佳的整体性能。使用最优算法重建后，训练集和测试集的AUC分别为0.85和0.77，准确率分别为78.8%和71.6%，召回率分别为75.6%和69.2%，DCA曲线均显示出更高的净收益。Shapley加性解释（SHAP值）和变量的优先级排序显示，涉及4个或更多器官的多器官功能障碍综合征是最重要的特征变量，其次是急性生理学和慢性健康状况评分Ⅱ和序贯器官衰竭评估（SOFA）评分。

通过机器学习建立模型可以准确预测脓毒性休克患者住院28 d死亡风险，改善其风险分层，从而指导临床医生采取适当的干预措施。

To establish a prediction model for 28-day in-hospital mortality of patients with septic shock using machine learning (ML) and verify its effectiveness.

The clinical data of patients with septic shock admitted to Suzhou Ninth Hospital Affiliated to Soochow University from January 2019 to June 2024 were analyzed, retrospectively. Patients were divided into death group (104 cases) and survival group (161 cases) based on their prognosis at 28 days after admission. The key variables associated with 28-day in-hospital mortality were screened by the least absolute shrinkage and selection operator (LASSO) algorithm. The entire dataset was randomly divided into training set and test set with a ratio of 3∶1. By comparing eight different ML algorithms, the optimal algorithm was selected to construct the prediction model. The model performance were evaluated by area under the receiver operating characteristic (ROC) curve (AUC) and clinical decision curve analysis (DCA). The model was interpreted by Shapley additive exPlanations (SHAP) algorithm.

Total of 265 patients with septic shock were enrolled. Forty-seven variables were initially included based on clinical preliminary screening; nine key variables were selected by the LASSO model for subsequent model development. Among the eight models constructed, the radial basis function-support vector machine (RSVM) model demonstrated the best overall performance. After optimization by this algorithm, the AUCs were 0.85 for the training set and 0.77 for the test set; the accuracies were 78.8% and 71.6%, respectively; the recall rates were 75.6% and 69.2%, respectively; and the DCA curves indicated higher net benefits. SHAP value analysis revealed that septic shock complicated with multiple organ dysfunction syndrome (involving four or more organs) was the most important predictive variable, followed by the acute physiology and chronic health evaluation Ⅱ (APACHE Ⅱ) score and the sequential organ failure assessment (SOFA) score.

Machine learning could be used to build an accurate prediction model for the 28-day in-hospital mortality risk of patients with septic shock. This model could improve the risk stratification and may guide clinicians in implementing appropriate interventions.