Predicting Acute Kidney Injury via Interpretable Ensemble Learning and Attention Weighted Convoutional-Recurrent Neural Networks
Predicting Acute Kidney Injury via Interpretable Ensemble Learning and Attention Weighted Convoutional-Recurrent Neural Networks
复制标题
DOI：
10.1109/ciss50987.2021.9400242
复制
发表时间：
2021-03
期刊：
影响因子：
--
通讯作者：
Yu-Chung Peng;N. S. D'Souza;Brian Bush;Charles H. Brown;A. Venkataraman
中科院分区：
文献类型：
--
作者：
Yu-Chung Peng;N. S. D'Souza;Brian Bush;Charles H. Brown;A. Venkataraman
急性肾损伤（AKI）是最常见的术后并发症之一，与短期和长期死亡率相关。改善AKI的预测是至关重要的，可以帮助临床医生预防和减轻其不良反应。在本文中，我们探索使用机器学习方法来预测术后AKI。我们的分析集中在基于集成的随机森林（RF）分类器上，该分类器在静态临床变量上运行，以及一种新的深度学习架构，该架构将术中时间序列数据与静态变量结合在一起。该体系结构使用双注意机制来选择与AKI预测相关的特征和时间间隔。我们在公开可用的VitalDB数据库上评估了3,640例接受非心脏手术的患者。RF优于AKI文献中现有的机器学习分类器（AUROC: 0.86, AUPRC: 0.54）。此外，射频识别了一组强大的术前变量，可以在简单的血液检查中筛选。虽然深度学习模型的性能略低（AUROC: 0.84, AUPRC: 0.44），但注意权值提供了重要的术中信息，临床医生可以在手术过程中对其进行监测。综上所述，我们的研究结果突出了机器学习在AKI预测方面的前景，并朝着开发临床可翻译模型迈出了第一步。
Acute Kidney Injury (AKI) is one of the most frequent postoperative complications and is associated with both short- and long-term mortality. Improved prediction of AKI is crucial and may help clinicians prevent and mitigate its adverse effects. In this paper, we explore the use of machine learning methods to predict postoperative AKI. Our analysis centers on the ensemble-based random forest (RF) classifier, which operates on static clinical variables, and a novel deep learning architecture that incorporates intraoperative time series data along with the static variables. The architecture uses a dual-attention mechanism to select both features and time intervals relevant for AKI prediction. We evaluate our models on the publicly available VitalDB database of 3,640 patients who underwent non-cardiac surgery. The RF outperformed existing machine learning classifiers in the AKI literature (AUROC: 0.86, AUPRC: 0.54). In addition, the RF identified a robust set of preoperative variables that can be screened in a simple blood test. While the deep learning model achieved slightly lower performance (AUROC: 0.84, AUPRC: 0.44), the attention weights provide important intraoperative information, which can be monitored by clinicians during surgery. Taken together, our results highlight the promise of machine learning for AKI prediction and take the first steps towards developing clinically translatable models.