通讯作者: 谢健, xiejian0518@sina.com

DOI：10.12201/bmr.202605.00087

A Prognostic Prediction Model for Elderly Sepsis Based on Transformer and Explainable Artificial Intelligence: A Multicenter Study

• 摘要：目的/意义　 老年脓毒症预后异质性大，现有预测工具在精准性与可解释性方面存在不足。本研究旨在构建并验证一个面向老年脓毒症患者的可解释的预后预测模型。方法/过程　回顾性分析eICU‐CRD（训练集）与MIMIC‐IV（外部验证集）数据库中符合Sepsis‐3.0标准的老年患者（≥60岁），提取13项核心炎症指标，基于Transformer架构构建可解释脓毒症预后预测模型（ESPT），并利用SHAP可解释框架进行全局特征重要性排序及个体预测的可视化解析。结果/结论　 ESPT模型在训练集的AUC为0.8589，在外部验证集中为0.8147，性能优于随机森林、XGBoost等传统机器学习模型。SHAP分析识别出血糖、血红蛋白和血细胞比容为关键的预后驱动因子。该模型可为老年脓毒症的风险分层与临床决策提供有效的辅助工具。

  关键词： 脓毒症; 预后模型; 可解释人工智能; 炎症生物标志物; 多中心研究

   

  Abstract: Purpose/Significance The prognosis of elderly sepsis is highly heterogeneous, and existing prediction tools lack both accuracy and interpretability. This study aims to develop and validate an interpretable prognostic prediction model specifically for elderly patients with sepsis. Methods/Process Elderly patients (≥60 years) meeting the Sepsis?3.0 criteria were retrospectively enrolled from the eICU?CRD (training set) and MIMIC?IV (external validation set) databases. Thirteen core inflammatory indicators were extracted, and an explainable sepsis prognosis transformer (ESPT) model was constructed based on the Transformer architecture. The SHAP framework was employed for global feature importance ranking and visual interpretation of individual predictions. Results/Conclusion The ESPT model achieved an area under the curve (AUC) of 0.8589 in the training set and 0.8147 in the external validation set, outperforming traditional machine learning models such as Random Forest and XGBoost. SHAP analysis identified blood glucose, hemoglobin, and hematocrit as key prognostic drivers. The ESPT model can serve as an effective auxiliary tool for risk stratification and clinical decision making in elderly patients with sepsis.

  Key words: Sepsis; Prognostic model; Explainable artificial intelligence; Inflammatory biomarkers; Multicenter study

  提交时间：2026-05-25

  版权声明：作者本人独立拥有该论文的版权，预印本系统仅拥有论文的永久保存权利。任何人未经允许不得重复使用。
• html

• 图表

• 丁荔. 镁离子水平与脓毒症预后的研究现状. 2025. doi: 10.12201/bmr.202509.00008

  刘淑月, 秦超. 肠道微生态对老年脓毒症预后影响的研究进展. 2025. doi: 10.12201/bmr.202501.00041

  向艾宁堃, 田靖雪, 胡德华, 刘海霞. 生成式人工智能对于老年糖尿病应答效能的比较研究. 2025. doi: 10.12201/bmr.202503.00019

  钱新阳, 王育民. pSOFA评分联合AGI分级评估脓毒症患儿预后的价值. 2024. doi: 10.12201/bmr.202411.00002

  马傲, 葛小玲. 人工智能大模型在医疗健康领域应用的研究. 2024. doi: 10.12201/bmr.202408.00039

  周澎洋, 刘莉莉, 陈思萱, 张宝林, 蒲旭虹. 基于人工智能大模型的中医电子病历内涵质控系统构建与实践. 2026. doi: 10.12201/bmr.202605.00088

  顾洁, 邢清敏, 肖涛, 李明. 乳酸/白蛋白比值联合APACHE II评分在脓毒症患者预后评估中的价值. 2026. doi: 10.12201/bmr.202603.00104

  哈力木拉提·买买提, 马翔. “人工智能+医疗卫生”背景下心音与心血管疾病的研究进展. 2026. doi: 10.12201/bmr.202601.00055

  杨晨, 余晓莹. 基于BERTopic的医疗人工智能公众舆论研究. 2026. doi: 10.12201/bmr.202609.00006

  郑琰莉, 韩福海, 李舒玉, 苏文星. 人工智能大模型在医疗领域的应用现状与前景展望. 2023. doi: 10.12201/bmr.202312.00027

• 序号	提交日期	编号	操作
  1	2025-11-29	10.12201/bmr.202605.00087V1	下载

• 公开评论  匿名评论  仅发给作者 提交评论