Department of Biomedical Data Intelligence, Graduate School of Medicine, Kyoto University, 54 Shogoin Kawahara-cho, Sakyo-ku, Kyoto 606-8507, Japan; Department of Nephrology, Graduate School of Medicine, Kyoto University, Japan.
Department of Biomedical Data Intelligence, Graduate School of Medicine, Kyoto University, 54 Shogoin Kawahara-cho, Sakyo-ku, Kyoto 606-8507, Japan.
Comput Methods Programs Biomed. 2021 Jul;206:106129. doi: 10.1016/j.cmpb.2021.106129. Epub 2021 Apr 27.
Acute kidney injury (AKI) occurs frequently in in-hospital patients, especially in the intensive care unit (ICU), due to various etiologies including septic shock. It is clinically important to identify high-risk patients at an early stage and perform the appropriate intervention.
We proposed a system to predict AKI using one-dimensional convolutional neural networks (1D-CNN) with the real-time calculation of the probability of developing AKI, along with the visualization of the rationale behind prediction using score-weighted class activation mapping and guided backpropagation. The system was applied to predicting developing AKI based on the KDIGO guideline in time windows of 24 to 48 h using data of 0 to 24 h after admission to ICU.
The comparison result of multiple algorithms modeling time series data indicated that the proposed 1D-CNN model achieved higher performance compared to the other models, with the mean area under the receiver operating characteristic curve of 0.742 ± 0.010 for predicting stage 1, and 0.844 ± 0.029 for stage 2 AKI using the input of the vital signs, the demographic information, and serum creatinine values. The visualization results suggested the reasonable interpretation that time points with higher respiratory rate, lower blood pressure, as well as lower SpO2, had higher attention in terms of predicting AKI, and thus important for prediction.
We presumed the proposed system's potential usefulness as it could be applied and transferred to almost any ICU setting that stored the time series data corresponding to vital signs.
急性肾损伤（AKI）在住院患者中很常见，尤其是在重症监护病房（ICU），其病因包括感染性休克。早期识别高危患者并进行适当干预具有重要的临床意义。
我们提出了一种使用一维卷积神经网络（1D-CNN）的系统，该系统可以实时计算发生 AKI 的概率，并使用基于得分加权类激活映射和引导反向传播的预测推理可视化来预测 AKI。该系统应用于根据 KDIGO 指南预测 ICU 入住后 0 至 24 小时内 24 至 48 小时的 AKI 发展，基于 ICU 入住后 0 至 24 小时的时间窗口。
对多种算法的时间序列数据建模结果进行比较表明，与其他模型相比，所提出的 1D-CNN 模型具有更高的性能，其预测 1 期和 2 期 AKI 的平均接收者操作特征曲线下面积分别为 0.742±0.010 和 0.844±0.029，输入包括生命体征、人口统计学信息和血清肌酐值。可视化结果表明，呼吸频率较高、血压较低、SpO2 较低的时间点在预测 AKI 时受到更高的关注，因此对预测很重要。
我们推测该系统具有潜在的应用价值，因为它可以应用于几乎任何存储与生命体征相对应的时间序列数据的 ICU 环境，并可进行转移。
