دانلود رایگان ترجمه مقاله سیستم پیشنهادی مبنی بر نمونه تصویر داده کاوی پزشکی (نشریه الزویر 2017)

سیستم پیشنهادی مبنی بر نمونه تصویر داده کاوی پزشکی

Visual instance-based recommendation system for medical data mining

Abstract

This paper presents an instance-based algorithm allowing exploration of large medical dataset by making pairwise connection between patients. In our metric-free method, each individual in a dataset ranks every member of the dataset. By aggregating these ranks, it is then possible to visualize data according to typical individuals representing subsets of closely-related patients. The paper also describes a visualization tool allowing exploration of a database of diabetic patients. This prototype of a recommendation system implements the aforementioned algorithm to enrich data, structure patients, create associations between individuals and provide recommendations.

1 Introduction

As electronic health records and wearable sensors become more widespread, medical datasets tend to be larger and call for specific methods of exploration. These datasets come with inherent problems : they contain high-dimensional data 1 which can be heterogeneous and unstructured, often including missing values. With such amount of available data, doctors need assistance to find relevant patients. They rely on recommendation systems to browse 2,3, explore and manipulate records, allowing them to find links between similar patients. Despite machine learning applications to medical data showing promising results 4, use of traditional approaches does not exactly fit the clinical context 5. These approaches tend to generalize and exclude outliers, algorithms need to be trained and most solutions prove to be “black boxes” that reduce interpretability6,7. While they perform well when trying to predict diseases, those algorithms are not suitable when trying to understand the same diseases and study atypical patients without heavily relying on domain expertise. The purpose of this paper is to provide a method able to structure elements of a dataset by creating associations between them. The resulting structure enriches data with measures of representativeness ,and is a way to visualize typical individuals or similar patients in a medical database. The proposed algorithm is instance-based, as to avoid overgeneralization. It is also a mean of simulating the reasoning of doctors faced with new patients : they match new individuals with past cases they encountered. Another advantage of this method is the ease with which a user can express constraints on the association rules between elements, making it a customizable tool. The main component of our method is an election algorithm where individuals attribute a score to every member of the dataset, in order to determine the most representative individuals across the population (this concept is close to class prototypes, but does not need any prior clustering step on the dataset 9). By using different functions to calculate the score of each individual, the output structure can easily be modified. This contribution is divided into the following sections : first we share some background related to our work, the problems we tackle and the data we use to illustrate this article. The second section focuses on concepts and algorithms that enrich data and create recommendations. The following section describes in details the visualization and exploration tool we built using the aforementioned algorithms. This paper ends with a discussion and presentation of future related works.