HMV: A medical decision support framework using multi-layer classifiers for disease prediction

Highlights

• Extensive research has been conducted on disease prediction.

• However, there is no agreement on which classifier produces the best results.

• An optimal combination of classifiers is presented with multi-layer classification.

• The ensemble approach uses bagging with multi-objective optimized weighted.

• Comparison with existing techniques shows superiority of our ensemble.

Abstract

Decision support is a crucial function for decision makers in many industries. Typically, Decision Support Systems (DSS) help decision-makers to gather and interpret information and build a foundation for decision-making. Medical Decision Support Systems (MDSS) play an increasingly important role in medical practice. By assisting doctors with making clinical decisions, DSS are expected to improve the quality of medical care. Conventional clinical decision support systems are based on individual classifiers or a simple combination of these classifiers which tend to show moderate performance. In this research, a multi-layer classifier ensemble framework is proposed based on the optimal combination of heterogeneous classifiers. The proposed model named “HMV” overcomes the limitations of conventional performance bottlenecks by utilizing an ensemble of seven heterogeneous classifiers. The framework is evaluated on two different heart disease datasets, two breast cancer datasets, two diabetes datasets, two liver disease datasets, one Parkinson's disease dataset and one hepatitis dataset obtained from public repositories. Effectiveness of the proposed ensemble is investigated by comparison of results with several well-known classifiers as well as ensemble techniques. The experimental evaluation shows that the proposed framework dealt with all types of attributes and achieved high diagnosis accuracy. A case study is also presented based on a real time medical dataset in order to show the high performance and effectiveness of the proposed model.

Introduction

Data mining in medical domain, is a process of discovering hidden patterns and information from large medical datasets; analyzes them and uses them for disease prediction [1]. The basic goal of data mining process is to extract hidden information from medical datasets and transform it into an understandable structure for future use [2]. A large number of predictive models can be developed from data mining techniques which enable classification and prediction tasks. After discovering knowledge from data, learning phase starts; where a scientific model is built. This learning method evolves the concept of machine learning and can be formally defined as “the complex computation process of automatic pattern recognition and intelligent decision making based on training sample data” [3]. The machine learning classifiers are further categorized into supervised learning and unsupervised learning depending on the availability of data. In supervised learning, labeled training data is available and a learning model is trained. Some examples include Artificial Neural Network (ANN), Support Vector Machine (SVM), and Decision Trees (DT). In unsupervised learning, there is no class label field in sample data. Examples include, K-mean clustering and Self-Organization Map (SOM). An ensemble approach performs better than individual machine learning techniques by combining the results of individual classifiers [4], [5]. There are multiple techniques that can be utilized for constructing the ensemble model and each result in different diagnosis accuracy. Most common ensemble approaches are bagging [6], boosting [7] and stacking [8].

Significant amount of work has already been done on disease classification and prediction. However, there is no single methodology which shows highest performance for all datasets or diseases, while one classifier shows good performance in a given dataset, another approach outperforms the others for other dataset or disease. The proposed research focuses on a novel combination of heterogeneous classifiers for disease classification and prediction, thus overcoming the limitations of individual classifiers. The novel combination of heterogeneous classifiers is presented which is Naïve Bayes, Linear Regression, Quadratic Discriminant Analysis, K-Nearest Neighbor, Support Vector Machine, Decision tree using Information Gain and Decision tree using the Gini Index. The multiple classifiers are used at multiple layers to further enhance disease prediction accuracy. An application has also been developed for disease prediction. It is based on the proposed HMV ensemble framework. The proposed application can help both doctors and patients in terms of data management and disease prediction.

The rest of the paper is organized as follows: Section 2 relates to literature review. Section 3 presents the proposed ensemble framework. Section 4 provides the results and discussion from the experiments carried out. Section 5 provides a case study about proposed ensemble model, whereas Section 6 is related to the discussion. Medical application for disease diagnosis is detailed in Section 7 and finally the conclusion is provided in Section 8.