In many medical subfields, there is a call for greater interpretability in the machine learning systems used for clinical work. In this paper, we design an interpretable deep learning model to predict the presence of 6 types of brainwave patterns (Seizure, LPD, GPD, LRDA, GRDA, other) commonly encountered in ICU EEG monitoring. Each prediction is accompanied by a high-quality explanation delivered with the assistance of a specialized user interface. This novel model architecture learns a set of prototypical examples (``prototypes'') and makes decisions by comparing a new EEG segment to these prototypes. These prototypes are either single-class (affiliated with only one class) or dual-class (affiliated with two classes). We present three main ways of interpreting the model: 1) Using global-structure preserving methods, we map the 1275-dimensional cEEG latent features to a 2D space to visualize the ictal-interictal-injury continuum and gain insight into its high-dimensional structure. 2) Predictions are made using case-based reasoning, inherently providing explanations of the form ``this EEG looks like that EEG.'' 3) We map the model decisions to a 2D space, allowing a user to see how the current sample prediction compares to the distribution of predictions made by the model. Our model performs better than the corresponding uninterpretable (black box) model with $p<0.01$ for discriminatory performance metrics AUROC (area under the receiver operating characteristic curve) and AUPRC (area under the precision-recall curve), as well as for task-specific interpretability metrics. We provide videos of the user interface exploring the 2D embedded space, providing the first global overview of the structure of ictal-interictal-injury continuum brainwave patterns. Our interpretable model and specialized user interface can act as a reference for practitioners who work with cEEG patterns.
翻译:在许多医疗领域中,要求临床工作中使用的机器学习系统更容易被理解。本文中,我们设计了一个可解释的深度学习模型来预测 ICU 脑电监测中常见的 6 种脑波模式(癫痫、LPD、GPD、LRDA、GRDA、other)是否存在,并借助用户界面提供高质量的解释。这种新型模型体系结构通过学习一组样本原型(“原型”)进行决策,并通过将新的 EEG 段与这些原型进行比较来实现。这些原型要么是单类(仅与一个类相关联),要么是双类(与两个类相关联)。我们提供了三种主要的解释模型的方法:1)使用全局结构保持方法,将 1275 维的 cEEG 潜在特征映射到 2D 空间中,以可视化癫痫-间歇性-损伤连续电势的连续谱,并深入了解其高维结构。2)利用基于案例的推理进行预测,从根本上提供形式为“这个 EEG 看起来像那个 EEG”的解释。3)我们将模型决策映射到 2D 空间中,允许用户看到当前样品预测与模型所做预测分布的比较。我们的模型在识别性能指标 AUROC(接收器操作特征曲线下的面积)和 AUPRC(精确率-召回率曲线下的面积)以及任务特定的可解释性指标方面表现优于相应的不可解释(黑盒)模型,具有 $p<0.01$的统计学意义。我们提供了使用 2D 嵌入空间探索用户界面的视频,提供了癫痫-间歇性-损伤连续电位脑电图模式结构的第一个全局概述。我们的可解释的模型和专用用户界面可以作为与 cEEG 模式工作的从业者的参考。