Early Identification of Acute Kidney Injury Using Deep Recurrent Neural Nets, Presented with Probable Etiology

点击翻译按钮获取中文摘要

Abstract Significance: In this SBIR project we propose to develop Previse, a novel, software-based clinical decision support (CDS) system for predicting acute kidney injury (AKI), and attributing AKI to one of several causal mechanisms (etiologies). Previse will use machine learning methods and information drawn from the electronic health record (EHR) to identify the early signs of acute kidney injury. By doing so before the clinical syndrome of AKI is fully developed, Previse will give clinicians the time to intervene with the goals of preventing further kidney damage, and decreasing the sequelae of AKI. Combining this prediction module with a second module that suggests the underlying causes responsible for an incipient or full AKI, Previse will enable clinicians to make earlier and better-informed treatment decisions for AKI patients. Research Question: Can a machine- learning-based CDS predict the development and progression of AKI in hospitalized patients 72 hours in advance of KDIGO stage 2 or 3, with performance providing an area under the receiver operating characteristic curve (AUROC) of at least 0.85? Is it possible to use a Bayesian model to infer the cause of AKI with high accuracy (AUROC ≥ 0.75)? Prior work: We have developed a prototype version of the Previse system which predicts AKI up to 72 hours in advance of KDIGO stage 2 or 3 criteria, with an AUROC near 0.70. We have previously developed machine-learning-based predictive tools for sepsis, in-hospital mortality, and other adverse patient events with performance significantly improved over commonly used rules-based scoring systems. Specific Aims: To predict the onset of chart-abstracted KDIGO stage 2 or 3 AKI in retrospective data, 72 hours in advance (Aim 1); to use data drawn from the EHR to identify the cause of AKI at time of onset with high accuracy, and to present this causal inference, its likelihood, and relevant evidence supporting it in a human-interpretable fashion (Aim 2). Methods: We will predict the onset of AKI using a deep, recurrent neural network (RNN). This expressive, nonlinear classifier will incorporate time-series information in the qualitative portions of the EHR and will also incorporate features derived from text components, such as radiology reports. Labeling AUROC of 0.85 or higher at 72 hours pre-KDIGO AKI will constitute success in Aim 1. In Aim 2, we will train a dynamic Bayesian network to identify the cause of AKI. We will train this system using semi-supervised methods, where the causes of a set of AKI examples will be hand-annotated by clinician experts; these examples will be split into two groups, with some used for training and the remainder for testing. Aim 2 will be successful if this training results in etiology identification accuracy of at least 0.75 in the test set. Future Directions: Following the proposed work, the combined Previse system will be deployed for prospective studies at partner hospitals.