Ellenius Johan, Groth Torgny
Department of LIME (Learning, Informatics, Management and Ethics), Karolinska Institutet, 171 77 Stockholm, Sweden.
Artif Intell Med. 2008 Mar;42(3):189-98. doi: 10.1016/j.artmed.2007.10.002.
A common objection to using artificial neural networks in clinical decision support systems is that the reasoning behind diagnostic indications cannot be sufficiently well explained. This paper presents a method for visualizing diagnostic indications generated from an artificial neural network-based decision support algorithm (ANN-algorithm) in conditions developing over time.
The main idea behind the method is first to calculate and graphically present the decision regions corresponding to the diagnostic indications given as output from the ANN-algorithm, in the space of two selected, clinically established 'display variables'. Secondly, the trajectory of time series measurement results of these, often biochemical markers, together with the respective 95% confidence intervals are superimposed on the decision regions. This will permit a nurse or clinician to grasp the diagnostic indication graphically at a glance. The indication is further presented in relation to clinical variables that the clinician is already familiar with, thus providing a sort of explanation. The predictive value of the indication is expressed by the proximity of the measurement result to the decision boundary, separating the decision regions, and by a numerically calculated individualized predictive value.
The method is illustrated as applied to a previously published ANN-algorithm for the early ruling-in and ruling-out of acute myocardial infarction, using monitoring of measurement results of myoglobin and troponin-I in plasma.
The method is appropriate when there is a limited number of clinically established variables, i.e. variables which the clinician is used to taking into account in clinical reasoning.
在临床决策支持系统中使用人工神经网络存在一个常见的反对意见，即诊断指征背后的推理无法得到充分解释。本文提出了一种方法，用于可视化基于人工神经网络的决策支持算法（ANN算法）在随时间发展的病情中生成的诊断指征。
该方法背后的主要思想是，首先在两个选定的、临床确定的“显示变量”空间中，计算并以图形方式呈现与ANN算法输出的诊断指征相对应的决策区域。其次，将这些通常为生化标志物的时间序列测量结果的轨迹及其各自的95%置信区间叠加在决策区域上。这将使护士或临床医生能够一眼直观地理解诊断指征。该指征还与临床医生已经熟悉的临床变量相关呈现，从而提供一种解释。指征的预测价值通过测量结果与分隔决策区域的决策边界的接近程度以及通过数值计算的个体化预测值来表示。
该方法应用于先前发表的用于急性心肌梗死早期排除和确诊的ANN算法，以血浆中肌红蛋白和肌钙蛋白I的测量结果监测为例进行说明。
当临床确定的变量数量有限时，即临床医生在临床推理中习惯考虑的变量数量有限时，该方法是合适的。
