南方医科大学学报

南方医科大学学报 ›› 2026, Vol. 46 ›› Issue (6): 1423-1433.doi: 10.12122/j.issn.1673-4254.2026.06.22

• • 上一篇    

基于Transformer的影像与检验数据多模态预测模型在医院获得性感染的应用研究

张明旭1,2(), 吴迪恒1,2, 林嘉泽1, 吕祎君2, 周凌宏1(), 甄鑫1(), 周浩3(), 秦耿耿1,2()   

  1. 1.南方医科大学生物医学工程学院，广东 广州 510515
    2.南方医科大学南方医院影像诊断科，广东 广州 510515
    3.南方医科大学珠江医院感染科，广东 广州 510280
  • 收稿日期:2025-09-08 出版日期:2026-06-20 发布日期:2026-06-24
  • 通讯作者: 周凌宏,甄鑫,周浩,秦耿耿 E-mail:826135508@qq.com;smart@smu.edu.cn;xinzhen@smu.edu.cn;630304495@qq.com;zealotq@smu.edu.cn
  • 作者简介:张明旭，在读硕士研究生，E-mail: 826135508@qq.com
  • 基金资助:
    国家自然科学基金(3257071033);广东省自然科学基金(2024A1515011520);广东省基础与应用基础研究基金(2024A1515012100)

A Transformer-based multimodal model for predicting hospital-acquired infections using imaging and clinical laboratory data

Mingxu ZHANG1,2(), Diheng WU1,2, Jiaze LIN1, Yijun LÜ2, Linghong ZHOU1(), Xin ZHEN1(), Hao ZHOU3(), Genggeng QIN1,2()   

  1. 1.School of Biomedical Engineering, Southern Medical University, Guangzhou 510515, China
    2.Department of Radiologic Diagnosis, Nanfang Hospital, Southern Medical University, Guangzhou 510515, China
    3.Department of Infectious Diseases, Zhujiang Hospital, Southern Medical University, Guangzhou 510280, China
  • Received:2025-09-08 Online:2026-06-20 Published:2026-06-24
  • Contact: Linghong ZHOU, Xin ZHEN, Hao ZHOU, Genggeng QIN E-mail:826135508@qq.com;smart@smu.edu.cn;xinzhen@smu.edu.cn;630304495@qq.com;zealotq@smu.edu.cn
  • Supported by:
    National Natural Science Foundation of China(3257071033)

RichHTML

32

PDF

115

摘要：

目的 构建一个基于Transformer的多模态数据编码模型用于预测医院获得性感染。 方法 收集公开医学数据库MIMIC-IV中300 000条患者的实验室检验数据，同时在南方医院收集1172例患者的实验室检验数据和其中274例患者的X线胸片。使用本研究提出的基于Transformer的编码模型处理数据并接入机器学习分类器对医院获得性感染（HAI）进行预测。提取X线胸片影像组学特征和深度特征，采用特征融合算法将其与实验室检验数据融合后对呼吸机相关性肺炎（VAP）进行预测。通过准确率 （ACC）、AUC、灵敏度 （SEN） 和特异度 （SPE）评价模型性能，并将本研究所提出算法与传统机器学习分类算法进行定量比较，验证模型的有效性和可行性。 结果 结果显示本研究所构建的模型在内部验证中AUC可达0.989；在外部验证中，在预测HAI问题上的最优模型的AUC达到0.98，在融合影像特征后预测VAP任务中，最优模型的AUC达到0.93，均高于预设的基线模型性能。 结论 基于Transformer的实验室检验数据处理模型在预测HAI的问题中具有优秀的预测能力和较高的临床应用价值，并拓展性的对HAI的亚型进行了预测，取得了较好的结果。

关键词: 医院获得性感染, 呼吸机相关性肺炎, 实验室检验数据, 深度学习, 特征融合

Abstract:

Objective To construct a Transformer-based multimodal data encoding model for predicting hospital-acquired infections (HAI). Methods Laboratory test data of 300 000 patients were extracted from the publicly available MIMIC-IV database. The laboratory data of 1172 patients and chest X-ray images from 274 of these patients were collected from Nanfang Hospital. A novel Transformer-based encoding model was developed to process the data, which was then connected to a machine learning classifier for predicting HAI. The radiomic and deep learning features were extracted from the chest X-ray images for predicting ventilator-associated pneumonia (VAP). These imaging features were subsequently integrated with the laboratory test data using a feature fusion algorithm. The model performance was evaluated by assessing the accuracy, the area under the ROC curve (AUC), sensitivity, and specificity. The proposed algorithm was quantitatively compared against traditional machine learning classifiers to validate its effectiveness and feasibility. Results The results demonstrated that the model developed in this study achieved an AUC of 0.989 in the internal validation set. In the external validation set, the optimal model for predicting HAI attained an AUC of 0.98, and following the integration of imaging features, the optimal model reached an AUC of 0.93 in the VAP prediction task, demonstrating superior performance over the baseline models. Conclusion The Transformer-based model for processing laboratory test data has excellent predictive capability and good clinical applicability for HAI prediction with also good performance for predicting VAP.

Key words: hospital-acquired infections, ventilator-associated pneumonia, laboratory test data, deep learning, feature fusion