Feature subset selection by genetic algorithms and estimation of distribution algorithms

Abstract

The transjugular intrahepatic portosystemic shunt (TIPS) is an interventional treatment for cirrhotic patients with portal hypertension. In the light of our medical staff’s experience, the consequences of TIPS are not homogeneous for all the patients and a subgroup dies in the first 6 months after TIPS placement. Actually, there is no risk indicator to identify this subgroup of patients before treatment. An investigation for predicting the survival of cirrhotic patients treated with TIPS is carried out using a clinical database with 107 cases and 77 attributes. Four supervised machine learning classifiers are applied to discriminate between both subgroups of patients. The application of several feature subset selection (FSS) techniques has significantly improved the predictive accuracy of these classifiers and considerably reduced the amount of attributes in the classification models. Among FSS techniques, FSS–TREE, a new randomized algorithm inspired on the new EDA (estimation of distribution algorithm) paradigm has obtained the best average accuracy results for each classifier.

Introduction

Portal hypertension is a major complication of chronic liver disease. By definition, it is a pathological increase in the portal venous pressure which results in formation of porto-systemic collaterals that divert blood from the liver to the systemic circulation. This is caused by both an obstruction to outflow in the portal flow as well as an increased mesenteric flow. In the western world, cirrhosis of the liver accounts for approximately 90% of the patients.

Of the sequelae of portal hypertension (i.e. varices, encephalopathy, hypersplenism, ascites), bleeding from gastro-oesophageal varices is a significant cause of early mortality (approximately 30–50% at the first bleed) [4], [41].

Many efforts have been made over the past decades in the treatment of portal hypertension. This has resulted in an increasing number of randomized trials and publications but, unfortunately, therapeutic decision is not easy [10].

The transjugular intrahepatic portosystemic shunt (TIPS) is an interventional treatment resulting in decompression of the portal system by creation of a side-to-side portosystemic anastomosis. Since its introduction over 10 years ago [39], [40] and despite the large number of published studies, many questions remain unanswered. Currently, little is known about the effects of TIPS on the survival of the treated patients.

Our medical staff has found that a subgroup of patients dies in the first 6 months after a TIPS placement and the rest survive for longer periods. Actually there is no risk indicator to identify both subgroups of patients. We are equally interested in the detection of both subgroups, giving the same relevance to the reduction of both error types.

A period of 6 months is chosen as we think that beyond that period, factors such as stenosis of the shunt and possible variceal rebleeding as a consequence, would compound the analysis. Furthermore, a critical criteria for choosing this period is that the average waiting time on a list for a liver transplant at the University Clinic of Navarra is approximately 6 months. The only published study [31] to identify a subgroup of patients who die within a period after a TIPS placement fixes the length of this period to 3 months. However, we think that our specific conditions really suggest lengthening this period to 6 months.

Traditionally, Pugh’s modification of the Child–Turcotte score (referred to as the Child–Pugh score) has been used to assess risk in patients undergoing portosystemic shunt surgery [37]. Although it is a classic score to assess the level of seriousness of a patient’s liver disease, it has inherent problems when applied to patients undergoing TIPS and it cannot be used to predict which patients will die within a certain period of time and which patients will survive that period. The several difficulties and innacuracies in applying the Child–Pugh score to predict survival periods have been detailed by Conn [8].

In 1980s and 1990s, researchers in artificial intelligence have developed new machine learning methods that construct predictive models from data, obtaining promising results in several clinical areas [9], [18], [35]. As far we know, these kinds of techniques have never been applied to TIPS indication or contraindication. Thus, we assume that the prediction of patient survival within 6 months after elective TIPS is well-suited to supervised machine learning methods.

We have performed a prospective study, building up a database which includes a structured and standardized history and clinical examination. In the study reported here, we concentrate on the task of predicting survival within 6 months after a TIPS placement of hospitalized patients from their findings before a TIPS setting. For this purpose, in the first step, four well-known supervised machine learning methods with a long tradition on medical applications such as a Naive Bayes classifier, a decision-tree technique, a rule-learning procedure and a nearest neighbor method are applied.

However, the used database has a large set of measured findings¹ and some of them seem to be irrelevant or redundant. It is well known that the accuracy of supervised machine learning methods is not monotonic regarding the inclusion of features [26]: irrelevant or redundant attributes, depending on the specific characteristics of the classifier, may degrade the accuracy of the classification model. Ohmann et al. [35], in a problem of an acute abdominal pain diagnosis, note that the high dimensionality of their study database is the major problem in order to improve the predictive accuracy of their supervised classification models. In this sense, given the entire set of attributes, we aim to find the attribute subset with the best predictive accuracy for a certain classifier. This problem is known in the machine learning community as the feature subset selection (FSS) problem and it has been tackled with success in different medical areas [13], [21]. A reduction in the number of variables is of interest as classification models with a smaller number of variables may be more quickly and easily used by clinicians, as these models would require a lower data input [9]. Models with a relatively small number of variables may be more readily converted into paper-based models that could be used widely in current medical practice. Other interesting effects of the dimensionality reduction are the decrease in the cost of adquisition of the data and the rise in the interpretability and comprehensibility of the classification models.

Thus, in the second stage of the study, we apply two sequential and two genetic FSS techniques over the same survival predictive problem. We extend our comparison with the application of two new FSS procedures based on the new EDA (estimation of distribution algorithm) [33] paradigm.

Although the application of two new EDA-inspired FSS techniques refers to the specific medical problem, the proposed approach is general and can be used for other tasks where supervised machine learning algorithms face a high number of irrelevant and/or redundant features.

Costs of medical tests are not considered in the construction of classification models and predictive accuracy maximization is the principal goal of our research. As the cost of the TIPS placement is not insignificant, our study is developed to help physicians, counsel patients and their families before deciding to proceed with elective TIPS.

The paper is organized as follows. The study database is described in Section 2. The supervised classifiers included in the study and the FSS methods to improve their predictive accuracy are described in Section 3. Experimental results are presented in Section 4. The last section briefly summarizes the work and presents ways of future research in the field.