合并糖尿病脓毒症患者脓毒症相关心肌病的 早期风险预测：模型开发与验证

Early Risk Prediction of Sepsis-Induced Cardiomyopathy in Septic Patients with Diabetes Mellitus: Model Development and Validation

摘要: 目的：脓毒性心肌病(Sepsis-Induced Cardiomyopathy, SIC)是脓毒症极为重要的器官功能障碍表型之一，与患者的不良结局密切相关。由于合并糖尿病的脓毒症患者存在更为复杂的代谢与炎症病理背景，早期识别SIC对开展临床风险分层与优化治疗决策具有重大意义。本研究基于MIMIC-IV v2.2数据库，旨在构建并验证合并糖尿病的脓毒症患者发生SIC的临床预测模型，并将其以可视化的列线图(nomogram)形式进行呈现。方法：本研究采用回顾性队列研究设计。研究对象筛选自2008至2019年间首次入住贝斯以色列女执事医疗中心(BIDMC)重症监护室(ICU)且ICU停留时间 ≥ 1天的成人脓毒症、严重脓毒症或败血性休克合并糖尿病患者。排除标准包括：① 出院诊断包含任何其他心脏疾病或心肌病；② 年龄 < 18岁；③ 未行超声心动图检查。SIC定义为：符合脓毒症相关诊断标准且存在左心室收缩功能受损(LVEF < 50%，或超声提示全局左心室运动不足/全局左心室收缩功能障碍)，并利用ICD-9/ICD-10代码剔除合并或既往有心脏病史的患者。候选预测因子限定为患者入住ICU前24小时内的人口学特征、生命体征、实验室指标、既往合并症与支持治疗信息。连续型变量经1%~99% Winsorize截尾处理后，依据训练集参数进行Z-score标准化；分类变量则进行哑变量转换。总体数据按7:3的比例随机划分为训练集与内部验证集。研究采用10折交叉验证的LASSO回归技术进行核心变量筛选，并据此建立多因素Logistic回归预测模型与列线图。模型的判别能力采用受试者工作特征曲线下面积(AUC/C-index)进行评估，校准能力则通过校准曲线及Brier分数评价；同时，模型亦在外部数据集中进行了验证。为校正模型在训练集中可能存在的表观过拟合乐观偏倚，本研究进一步运用Bootstrap重抽样(B = 300)策略对模型性能进行了严格的内部校正。结果：最终共纳入2844例合并糖尿病的脓毒症ICU患者，其平均年龄为69.3 ± 12.6岁，男性占比57.7%，SIC发生率为28.8%。其中训练集包含1990例(事件率28.8%)，测试集包含854例(事件率28.8%)。LASSO回归最终成功筛选出24个候选预测因子(涵盖SOFA评分、体重、收缩压、平均动脉压等)，并据此构建了Logistic回归模型。该模型在训练集中的判别能力AUC为0.743，在内部验证集中的AUC为0.724；两集的Brier分数分别为0.173和0.179。经Bootstrap乐观偏倚校正后，训练集表观性能的乐观偏倚极小：AUC由0.723微调至0.719，Brier分数由0.178校正为0.179，校准截距为−0.019，校准斜率达到0.978。结论：本研究基于大型规范重症数据库MIMIC-IV v2.2构建的针对合并糖尿病脓毒症患者SIC发病风险的LASSO-Logistic预测模型，在严密的内部交叉验证测试中展现出了令人满意的疾病区分与判别性能，且其可视化的列线图能够极大辅助一线医生在早期迅速实现精准的高危筛查。

Abstract: Objective: Sepsis-induced cardiomyopathy (SIC) is a critical organ dysfunction phenotype in sepsis, closely associated with adverse outcomes. Patients with sepsis and pre-existing diabetes present a more complex metabolic-inflammatory milieu, making early identification of SIC crucial for risk stratification and therapeutic decision-making. Based on the MIMIC-IV v2.2 database, this study aimed to develop and validate a clinical prediction model for SIC occurrence in septic patients with diabetes, presented in the form of a nomogram. Methods: This retrospective cohort study included patients with sepsis, severe sepsis, or septic shock admitted to the intensive care unit (ICU) of Beth Israel Deaconess Medical Center (BIDMC) between 2008 and 2019. Adult patients with pre-existing diabetes, admitted to the ICU for the first time with an ICU stay ≥1 day, were enrolled. Exclusion criteria were: 1) discharge diagnosis of any other cardiac disease/cardiomyopathy; 2) age < 18 years; 3) absence of echocardiographic examination. SIC was defined as meeting sepsis-related diagnostic criteria along with left ventricular systolic dysfunction (LVEF < 50%, or unreported LVEF with echocardiography indicating “global left ventricular hypokinesis/systolic dysfunction”), while patients with pre-existing or concurrent cardiac diseases were excluded via ICD-9/ICD-10 codes. Candidate predictors included demographic characteristics, vital signs, laboratory tests, comorbidities, and supportive treatments within the first 24 hours after ICU admission. Continuous variables were Winsorized at the 1st~99th percentiles and then Z-score standardized using the training-set parameters; categorical variables were converted into dummy variables. The dataset was randomly split into a training set and an internal validation set at a 7:3 ratio. Predictor selection was performed using 10-fold cross-validated LASSO, followed by multivariable logistic regression based on the selected predictors, and a nomogram was constructed. Discrimination was assessed using the area under the receiver operating characteristic curve (AUC)/C-index, and calibration was evaluated using calibration curves and the Brier score. External validation was conducted in an independent cohort. To correct for optimism in apparent performance due to overfitting, internal validation was additionally performed using bootstrap optimism correction (B = 300). Results: A total of 2844 ICU patients with sepsis and concomitant diabetes were included. The mean age was 69.3 ± 12.6 years, 57.7% were male, and the incidence of sepsis-induced cardiomyopathy (SIC) was 28.8%. The cohort was split into a training set (n = 1990; event rate 28.8%) and an internal validation set (n = 854; event rate 28.8%). LASSO ultimately selected 24 predictors (including sofa_score, weight, sbp, mbp, resp_rate, temperature, spo2, chloride, potassium, creatinine, platelet, hemoglobin, lactate, po2, pco2, congestive_heart_failure, peripheral_vascular_disease, chronic_pulmonary_disease, mild_liver_disease, apsiii, gcs, ventilator_hours, and gender_male), based on which a logistic regression model was developed. The model showed good discrimination with an AUC of 0.743 in the training set and 0.724 in the internal validation set; the Brier scores were 0.173 and 0.179, respectively. After bootstrap optimism correction (B = 300), the optimism in apparent training performance was small: the AUC was corrected from 0.723 to 0.719, and the Brier score from 0.178 to 0.179; the calibration intercept was −0.019 and the calibration slope was 0.978. Conclusions: The LASSO-Logistic prediction model developed in this study, based on the large-scale standardized critical care database MIMIC-IV v2.2 for predicting the risk of SIC in septic patients with diabetes mellitus, demonstrated satisfactory discrimination and predictive performance during rigorous internal cross-validation. Furthermore, its visualized nomogram can significantly assist frontline physicians in rapidly and accurately screening high-risk patients at an early stage.