胰十二指肠切除术患者住院期间急性肾损伤 相关临床特征分析及风险识别： 基于MIMIC-IV数据库的回顾性研究

Clinical Characteristics and Exploratory Risk Identification for In-Hospital Acute Kidney Injury in Patients Undergoing Pancreaticoduodenectomy: A Retrospective Study Based on the MIMIC-IV Database

摘要: 目的：基于公共重症医学数据库MIMIC-IV 2.2，分析接受胰十二指肠切除术(pancreaticoduodenectomy, PD)患者住院期间急性肾损伤(acute kidney injury, AKI)记录状态的相关临床特征，并基于logistic回归结果进行探索性风险识别可视化展示。方法：纳入MIMIC-IV 2.2中接受PD的成年ICU患者，PD基于ICD-9/ICD-10-PCS手术编码识别，AKI基于住院期间记录的ICD-9/ICD-10诊断编码定义。提取人口学特征、合并症、严重程度评分、实验室指标及治疗相关变量。采用单因素logistic回归筛选与住院期间AKI记录状态相关的候选变量，随后采用多因素logistic回归分析独立相关因素。在多因素logistic回归基础上，探索性绘制列线图，并采用ROC曲线及AUROC描述当前样本中AKI与非AKI记录状态的表观区分能力。鉴于本研究为回顾性数据库分析，且部分变量可能与AKI在住院过程中同期发生，列线图和AUROC仅用于探索性可视化和状态识别描述，不作为经验证的前瞻性预测模型或临床风险分层工具解释。结果：共纳入PD患者154例，其中住院期间记录AKI 29例(18.8%)。单因素分析提示年龄、凝血功能指标(INR, PTT)、血钠、SOFA评分、SAPS II评分及脓毒症等与住院期间AKI记录状态存在关联。多因素logistic回归显示，血管活性药物使用(OR = 3.14, 95% CI: 1.25~7.88, P = 0.015)与脓毒症(OR = 3.27, 95% CI: 1.14~9.36, P = 0.027)为住院期间AKI记录状态的独立相关因素。探索性ROC曲线分析显示，该logistic回归模型在当前样本中的表观AUROC为0.8116。结论：在MIMIC-IV数据库接受PD的ICU患者中，脓毒症和血管活性药物使用与住院期间AKI记录状态独立相关，提示感染负荷、循环不稳定和重症治疗支持需求可能是该人群AKI相关临床表型的重要组成部分。探索性列线图和ROC曲线可直观展示相关变量组合对AKI记录状态的表观区分趋势；但由于本研究为回顾性数据库分析，AKI依据诊断编码定义，且未进行内部验证、校准度评价及外部验证，相关图形不能作为成熟前瞻性预测模型或临床决策工具使用。

Abstract: Objective: Based on the public critical care database MIMIC-IV version 2.2, this study aimed to analyze clinical characteristics associated with recorded in-hospital acute kidney injury (AKI) among ICU patients undergoing pancreaticoduodenectomy (PD) and to provide an exploratory visualization of AKI status identification based on logistic regression results. Methods: Adult ICU patients who underwent PD in MIMIC-IV 2.2 were included. PD was identified using ICD-9/ICD-10-PCS procedure codes, and AKI was defined according to ICD-9/ICD-10 diagnosis codes recorded during hospitalization. Demographic characteristics, comorbidities, severity scores, laboratory indicators, and treatment-related variables were extracted. Univariate logistic regression was used to screen candidate variables associated with recorded in-hospital AKI, followed by multivariable logistic regression to identify independent associated factors. Based on the multivariable logistic regression model, a nomogram was constructed exploratorily, and a receiver operating characteristic curve analysis with the area under the curve was used to describe the apparent discrimination between recorded AKI and non-AKI status in the current sample. Given the retrospective database-based design and the possibility that some variables occurred concurrently with AKI during hospitalization, the nomogram and AUROC were used only for exploratory visualization and status-identification description rather than as a validated prospective prediction model or clinical risk-stratification tool. Results: A total of 154 patients undergoing PD were included, of whom 29 patients had AKI recorded during hospitalization, with an incidence of 18.8%. Univariate analysis suggested that age, coagulation indicators including international normalized ratio (INR) and partial thromboplastin time (PTT), serum sodium, SOFA score, SAPS II score, and sepsis were associated with recorded in-hospital AKI. Multivariable logistic regression showed that vasopressor use (OR = 3.14, 95% CI: 1.25~7.88, P = 0.015) and sepsis (OR = 3.27, 95% CI: 1.14~9.36, P = 0.027) were independently associated with recorded in-hospital AKI. Exploratory ROC analysis showed an apparent AUROC of 0.8116 in the current sample. Conclusion: Among ICU patients undergoing PD in the MIMIC-IV database, sepsis and vasopressor use were independently associated with recorded in-hospital AKI, suggesting that infection burden, circulatory instability, and the need for critical care support may be important components of AKI-related clinical phenotypes in this population. The exploratory nomogram and ROC curve visually demonstrated the apparent discrimination of the combined variables for recorded AKI status; however, because of the retrospective database-based design, diagnosis-code-based outcome definition, and lack of internal validation, calibration assessment, and external validation, these graphics should not be interpreted as a mature prospective prediction model or clinical decision-support tool.