文章摘要

文章信息/Info

Title:: Research on risk prediction of acute kidney injury based on CTSA-Net

作者:: 张青松1; 陈春晓1 △; 陈利海2△; （1.南京航空航天大学生物医学工程系，南京 211106；2.南京市第一医院麻醉科，南京 210006）

Author(s):: ZHANG Qingsong1; CHEN Chunxiao1; CHEN Lihai2; （1.Department of Biomedical Engineering, Nanjing University of Aeronautics and Astronautics, Nanjing 211106, China; 2.Anesthesiology Department of Nanjing First Hospital, Nanjing 210006）

Keywords:: Acute kidney injury; Deep learning; Electronic health records; Attention; Convolutional neural network

摘要:: 针对过去对急性肾损伤（acute kidney injury, AKI）患者的识别存在临床时间序列数据未被充分利用、提前预测窗口较短及缺少连续预测研究等不足，本研究提出了一种卷积神经网络和两阶段交叉注意力的混合网络模型（CTSA-Net），实现对1期及以上AKI的每小时连续预测。CTSA-Net的注意力支路、CNN支路及特征融合模块可增强对时间序列数据的全局表示以及局部细节的感知能力，从而提高对AKI的连续预测性能。在AKI发生时、发生前24、48及72 h四个预测时间点，模型预测AKI的受试者工作特征曲线下面积分别为0.946、0.907、0.895和0.879，准确率-召回率曲线下面积分别为0.979、0.960、0.949和0.939。实验结果表明，CTSA-Net模型在多个预测时间点进行AKI预测的性能较好，可用于患者的实时监测，辅助医生进行临床决策。

Abstract:: Addressing limitations in prior research of acute kidney injury（AKI）, including underutilization of clinical time series data, short predictive windows, and a lack of continuous prediction studies, we proposed a hybrid network model called CTSA-Net, integrated convolutional neural networks (CNN) and a two-stage cross-attention mechanism. CTSA-Net’s attention pathway, CNN pathway, and feature fusion module enhanced global representation of time series data and perception of local details, thereby improved the continuous prediction performance for AKI. At four different prediction time points at AKI onset, 24, 48, and 72 h before AKI onset, the model achieved respective area under the receiver operated characteristic curve (AUC) values of 0.946, 0.907, 0.895, and 0.879, respectively. The area under the precision-recall curve (PR-AUC) values were 0.979, 0.960, 0.949, and 0.939, respectively. Experimental results indicate that the CTSA-Net model demonstrates robust performance in predicting AKI at multiple time points, making it suitable for real-time patient monitoring and assisting clinicians in making informed clinical decisions.