LIMICS, Université Paris 13, Sorbonne Universités, INSERM UMRS 1142, 93017 Bobigny, France.
School of Computing and Mathematics, Ulster University, United Kingdom.
Artif Intell Med. 2019 Mar;94:42-53. doi: 10.1016/j.artmed.2019.01.001. Epub 2019 Jan 14.
Case-Based Reasoning (CBR) is a form of analogical reasoning in which the solution for a (new) query case is determined using a database of previous known cases with their solutions. Cases similar to the query are retrieved from the database, and then their solutions are adapted to the query. In medicine, a case usually corresponds to a patient and the problem consists of classifying the patient in a class of diagnostic or therapy. Compared to "black box" algorithms such as deep learning, the responses of CBR systems can be justified easily using the similar cases as examples. However, this possibility is often under-exploited and the explanations provided by most CBR systems are limited to the display of the similar cases. In this paper, we propose a CBR method that can be both executed automatically as an algorithm and presented visually in a user interface for providing visual explanations or for visual reasoning. After retrieving similar cases, a visual interface displays quantitative and qualitative similarities between the query and the similar cases, so as one can easily classify the query through visual reasoning, in a fully explainable manner. It combines a quantitative approach (visualized by a scatter plot based on Multidimensional Scaling in polar coordinates, preserving distances involving the query) and a qualitative approach (set visualization using rainbow boxes). We applied this method to breast cancer management. We showed on three public datasets that our qualitative method has a classification accuracy comparable to k-Nearest Neighbors algorithms, but is better explainable. We also tested the proposed interface during a small user study. Finally, we apply the proposed approach to a real dataset in breast cancer. Medical experts found the visual approach interesting as it explains why cases are similar through the visualization of shared patient characteristics.
基于案例的推理 (CBR) 是一种类比推理形式，其中通过使用包含先前已知案例及其解决方案的数据库来确定（新）查询案例的解决方案。从数据库中检索与查询相似的案例，然后对其解决方案进行调整以适用于查询。在医学中，案例通常对应于患者，问题在于将患者分类为诊断或治疗的某一类。与深度学习等“黑盒”算法相比，CBR 系统的响应可以很容易地使用相似案例作为示例进行解释。然而，这种可能性经常没有得到充分利用，大多数 CBR 系统提供的解释仅限于显示相似案例。在本文中，我们提出了一种 CBR 方法，该方法既可以作为算法自动执行，也可以在用户界面中以可视化方式呈现，以提供视觉解释或进行视觉推理。在检索相似案例后，可视化界面显示查询和相似案例之间的定量和定性相似性，以便人们可以通过视觉推理轻松地对查询进行分类，并且以完全可解释的方式进行分类。它结合了定量方法（通过基于多维缩放的极坐标中的散点图可视化，保留涉及查询的距离）和定性方法（使用彩虹框进行集可视化）。我们将该方法应用于乳腺癌管理。我们在三个公共数据集上证明了我们的定性方法具有与 k-最近邻算法相当的分类准确性，但可解释性更好。我们还在一项小型用户研究中测试了所提出的界面。最后，我们将所提出的方法应用于乳腺癌的真实数据集。医学专家发现，这种视觉方法很有趣，因为它通过可视化共享患者特征来解释为什么案例相似。
