南方医科大学学报

南方医科大学学报 ›› 2026, Vol. 46 ›› Issue (6): 1339-1348.doi: 10.12122/j.issn.1673-4254.2026.06.14

• • 上一篇    

急性胰腺炎患者早期有创机械通气需求模型的构建与验证：基于可解释机器学习

吴细旋(), 沙桐(), 胡鸿彬, 孙毛毛, 吴洁(), 陈仲清()   

  1. 南方医科大学南方医院重症医学科，广东 广州 510515
  • 收稿日期:2026-02-05 出版日期:2026-06-20 发布日期:2026-06-24
  • 通讯作者: 吴洁,陈仲清 E-mail:48164849@qq.com;shatong_smu@163.com;wujie11@smu.edu.cn;zhongqingchen2008@163.com
  • 作者简介:吴细旋，在读硕士研究生，E-mail: 48164849@qq.com
    沙 桐，主治医师，E-mail: shatong_smu@163.com
    第一联系人：同等贡献作者
  • 基金资助:
    国家自然科学基金(82572463);国家自然科学基金(82402512);大学生创新创业训练计划项目(S202412121144)

Development and validation of an interpretable machine learning model for predicting reguirement of early invasive mechanical ventilation in patients with acute pancreatitis

Xixuan WU(), Tong SHA(), Hongbin HU, Maomao SUN, Jie WU(), Zhongqing CHEN()   

  1. Department of Critical Care Medicine, Nanfang Hospital, Southern Medical University, Guangzhou 510515, China
  • Received:2026-02-05 Online:2026-06-20 Published:2026-06-24
  • Contact: Jie WU, Zhongqing CHEN E-mail:48164849@qq.com;shatong_smu@163.com;wujie11@smu.edu.cn;zhongqingchen2008@163.com
  • Supported by:
    National Natural Science Foundation of China(82572463)

RichHTML

28

PDF

123

摘要：

目的 探讨急性胰腺炎患者早期呼吸衰竭的风险预测因子，并利用可解释的机器学习模型，预测ICU入院后72 h内是否需要启动有创机械通气。 方法 基于MIMIC-IV数据库纳入急性胰腺炎成人患者进行回顾性队列研究。排除入ICU时已接受有创机械通气者，主要结局为入院后72 h内启动有创机械通气。构建多种机器学习模型并进行内部验证，选取表现最佳模型进行外部验证（eICU-CRD数据库）。采用SHapley 加法解释（SHAP）方法评估模型可解释性。 结果 纳入MIMIC-IV数据库517例急性胰腺炎患者，其中361例患者用于训练队列，156例患者用于内部测试队列。纳入eICU-CRD 269例急性胰腺炎患者作为外部验证队列。早期呼吸衰竭的患者在ICU入院时即表现出更显著的器官功能衰竭，且随机森林模型在训练集的AUC为0.908，在验证集的AUC为0.733，在外部验证集的AUC为0.681。序贯器官衰竭评分、血管活性药物使用及急性肾损伤是预测早期呼吸衰竭的关键因素。 结论 本研究构建的可解释机器学习模型在内外部验证中均显示出一定的预测效能，可用于评估急性胰腺炎患者早期有创机械通气风险。模型关键变量主要反映多器官功能衰竭状态，与疾病病理生理过程一致。

关键词: 多脏器功能衰竭, 呼吸衰竭, 重症医学, 决策支持系统, 可解释人工智能

Abstract:

Objective To explore the risk factors for early respiratory failure in patients with acute pancreatitis (AP) and develop an interpretable machine learning model to predict the need for invasive mechanical ventilation (MV) within 72 h of ICU admission. Methods This retrospective cohort study was conducted using data of adult patients with acute pancreatitis from the MIMIC-IV database, excluding those receiving invasive MV at ICU admission. The primary outcome was initiation of invasive MV within 72 h after ICU admission. Multiple machine learning models were developed and internally validated, and the best-performing model was externally validated using the eICU-CRD database. Model interpretability was assessed using SHapley Additive exPlanations (SHAP). Results A total of 517 patients with acute pancreatitis from the MIMIC-IV database were included, with 361 assigned to the training cohort and 156 to the internal testing cohort; 269 patients from the eICU-CRD database were included as the external validation cohort. The patients who developed early respiratory failure exhibited more severe organ dysfunction at ICU admission. The random forest model achieved an AUC of 0.908 in the training cohort, 0.733 in the internal testing cohort, and 0.681 in the external validation cohort. SHAP analysis identified sequential organ failure assessment (SOFA) score, use of vasoactive agents, and acute kidney injury (AKI) as the most important predictive factors. Conclusion The interpretable machine learning model demonstrates moderate predictive performance in both the internal and external validation cohorts and can be used for early risk assessment of invasive MV in patients with acute pancreatitis. The key predictive variables primarily reflect multi-organ dysfunction consistent with the underlying pathophysiological mechanisms of acute pancreatitis.

Key words: multiple organ failure, respiratory failure, critical care, decision support systems, interpretable artificial intelligence