A hybrid intelligent system for medical data classification

https://doi.org/10.1016/j.eswa.2013.09.022Get rights and content

Highlights

  • A hybrid intelligent system is proposed for medical data classification tasks.

  • The proposed system is able to learn incrementally and explain its predictions.

  • Benchmark medical data sets are used to evaluate the effectiveness of the system.

  • The results ascertain the usefulness of the system for medical decision support.

  • The knowledge base is presented as a decision tree for interpretation by users.

Abstract

In this paper, a hybrid intelligent system that consists of the Fuzzy Min–Max neural network, the Classification and Regression Tree, and the Random Forest model is proposed, and its efficacy as a decision support tool for medical data classification is examined. The hybrid intelligent system aims to exploit the advantages of the constituent models and, at the same time, alleviate their limitations. It is able to learn incrementally from data samples (owing to Fuzzy Min–Max neural network), explain its predicted outputs (owing to the Classification and Regression Tree), and achieve high classification performances (owing to Random Forest). To evaluate the effectiveness of the hybrid intelligent system, three benchmark medical data sets, viz., Breast Cancer Wisconsin, Pima Indians Diabetes, and Liver Disorders from the UCI Repository of Machine Learning, are used for evaluation. A number of useful performance metrics in medical applications which include accuracy, sensitivity, specificity, as well as the area under the Receiver Operating Characteristic curve are computed. The results are analyzed and compared with those from other methods published in the literature. The experimental outcomes positively demonstrate that the hybrid intelligent system is effective in undertaking medical data classification tasks. More importantly, the hybrid intelligent system not only is able to produce good results but also to elucidate its knowledge base with a decision tree. As a result, domain users (i.e., medical practitioners) are able to comprehend the prediction given by the hybrid intelligent system; hence accepting its role as a useful medical decision support tool.

Introduction

Research in computerized intelligent systems for medical applications is an important and exciting domain. In general, a physician typically accumulates his/her knowledge based on patients’ symptoms and the confirmed diagnoses. In other words, prognostic relevance of symptoms towards certain diseases and diagnostic accuracy of a patient are highly dependent on a physician’s experience (Meesad & Yen, 2003). As medical knowledge and treatment therapy progress rapidly, e.g. the occurrence of new diseases and the availability of new drugs, it is challenging for a physician to keep up-to-date with all recent knowledge and development in clinical practice (Meesad & Yen, 2003). On the other hand, with the advent of computing technologies, it is now relatively easy to acquire and store a lot of information digitally, e.g. in dedicated databases of electronic patient records (Pavlopoulos & Delopoulos, 1999). As such, the deployment of computerized medical decision support systems becomes a viable approach to assisting physicians to swiftly and accurately diagnose patients (Chabat, Hansell, & Yang, 2000). Nevertheless, numerous issues have to be overcome before a useful medical decision support system can be developed and deployed, which include decision making in the presence of uncertainty and imprecision (Tsipouras, Voglis, & Fotiadis, 2007). While medical experts’ knowledge and experience is important, ranging from assessing a patient’s condition to making a diagnosis, advances in machine learning (Kwiatkowska, Atkins, Ayas, & Ryan, 2007) techniques have opened up the way for medical practitioners to exploit computerized intelligent systems for decision support in their workplace, e.g. surgical imagery and X-ray photography (Isola, Carvalho, & Tripathy, 2012). When treating a patient, a physician first needs to narrow down the suspected disease to the root cause (out of a list of probable causes with similar symptoms) using his/her knowledge and experience, and then confirms the diagnosis by performing a number of tests (Isola et al., 2012). Concomitantly, computerized intelligent systems can be useful in assisting the physician to arrive at an informed decision quickly, e.g. by learning from similar past cases in a large database of electronic patient records and inferring the diagnosis for the current patient with proper justifications. The advantages of using such intelligent systems include increasing diagnosis accuracy and, at the same time, reducing time and costs associated with patient treatment (Çomak, Polat, Güneş, & Arslan 2007).

Machine learning models have been developed to support various medical decision making tasks. As an example, intelligent classifiers have been used for prognosis, diagnosis, and screening of diabetes, breast cancer and Parkinsons disease (Luukka, 2011). A number of neural-fuzzy models have been used as classifiers for heart disease, because they are capable of learning from data samples (i.e., patient records) and generalizing beyond the training samples (Kahramanli & Allahverdi, 2009). These include fuzzy neural networks, fuzzy probabilistic neural networks, and fuzzy learning vector quantization networks (Sekar, Dong, Shi, & Hu, 2012). However, one key limitation of these models is the lack of ability to explain their predictions (Markowska-Kaczmar & Matkowski, 2006). This is the motivation of this research, whereby we attempt to devise a machine learning-based system that is able to reveal its reasoning in dealing with an input case, and to provide justification for its predictions.

In machine learning, neural networks have significant advantages for medical decision support applications (Downs, Harrison, Kennedy, & Cross, 1996). Compared with expert systems, neural networks avoid the time-consuming and demanding knowledge acquisition process by directly learning complex association between input symptoms and target diseases from data samples i.e., patient records (Hayes-Roth, Waterman, & Lenat, 1983). In addition to learning, neural networks possess other useful properties, which include handling incomplete or missing data as well as filtering noise, uncertainty or imprecision (Downs et al., 1996). In view of the salient features of neural networks, the Fuzzy Min–Max (FMM) neural network is investigated for developing a usable and useful medical decision support tool in this paper. In order to further strengthen the FMM network for medical applications, the Classification and Regression Tree (CART) and Random Forest (RF) models are incorporated to produce a hybrid intelligent system. The proposed hybrid model serves as an extension of our previous work (Seera, Lim, Ishak, & Singh, 2012), which was focused on an offline model (i.e. FMM-CART) for undertaking fault detection and diagnosis problems. Here, the motivation is to extend the hybrid model such that it is equipped with the necessary characteristics for undertaking medical decision support tasks. While CART has the advantage of rule extracting in the form of a tree structure, it is less flexible in performing incremental learning from data samples. While FMM has the advantage of one-pass training with incremental learning properties, it lacks the capability of producing rules to explain its predictions. On the other hand, RF has the benefit of forming an ensemble of CART whereby the best tree can be identified to produce high prediction accuracy. Therefore, the hybrid model, i.e. FMM-CART-RF, has three distinctive capabilities, viz, learning incrementally from data samples (owing to FMM), explaining its predicted outputs (owing to CART), and achieving high classification performances (owing to RF). This is the key contribution of this research.

From the perspective of decision making, we commonly seek a second opinion (and sometimes even more) before making important decisions, especially one that has medical implications (Polikar, 2006). Then, different opinions are weighed and combined based on a thought process before the final decision is made. In machine learning, such ensemble concept is utilized too in developing a highly accurate model. In particular, bagging algorithms are useful for constructing an ensemble of decision trees, and one such variant is RF. RF is an effective ensemble method for data mining (Zhang, Zulkernine, & Haque, 2008). It has shown good results in many applications, which include automatic intrusion detection systems (Zhang et al., 2008) and medical data classification (Wu, Ye, Liu, & Ng, 2012). RF has some important characteristics such as providing useful internal estimates of strength, correlation, and variable importance, while producing high accuracy in comparison with many standard classification models (Wu et al., 2012). As a result, RF is exploited in this paper to form a group of diverse CART models so that the best tree can be chosen for classification and rule extraction purposes.

The hybrid model proposed in this research has two important practical implications in the domain of medical decision support. Firstly, the ability to provide explanation and justification for the prediction is of paramount importance, in order to convince domain users (i.e., medical practitioners) with the outcome given by a computerized decision support system. This ability is essential in safety critical applications, such as medical diagnosis and prognosis, whereby domain users need to understand, and be convinced of, how the computerized system arrives at such a prediction (Economou, Goumas, & Spiropoulos, 1996). The elucidated rules in the form of a decision tree from the hybrid model is, therefore, important in practice, whereby the rules could serve as a source of second opinions in medical diagnostic situations (Kovalerchuk, Vityaev, & Ruiz, 2000). Secondly, accuracy of a decision support system is very crucial in medical applications. As stated in Luukka (2011), a high false negative rate of a screening system would increase the risk of patients by depriving them from getting the necessary medical attention, while a high false alarm rate would cause unnecessary worry and stress in patients as well as increase the demand on medical resources. One the other hand, as indicated in Kinney (2003), a decision support system with high specificity and variable sensitivity could save medical costs and improve scheduling of vestibular patients in an otolaryngology clinic. Besides that, Huang, Yang, King, and Lyu (2006) also recognized the usefulness of machine learning models in reducing cost and saving time for undertaking medical diagnostic tasks. As shown in the experimental study, the proposed hybrid model not only is able to achieve high accuracy, sensitivity, and specificity rates, but also to provide explanation for its predictions in the form of a decision tree; hence demonstrating its usefulness as a decision support system in practical environments.

The organization of this paper is as follows. Literature reviews related to intelligent models for medical applications as well as rule extraction methods are detailed in Section 2. In Section 3, the dynamics of the proposed FMM-CART-RF model is presented. The experimental study, results, and discussion using a number of benchmark medical data sets are presented in Section 4. Finally, conclusions and suggestions for further work are presented in Section 5.

Section snippets

Literature Review

The literature review presented in this section is divided into two main parts: (i) intelligent systems for medical applications; (ii) rule extraction methods the emphasis on medical applications. In machine learning, supervised learning is a commonly used method for tackling medical problems. The task of identifying the smallest sets of genes and constructing a highly accurate classification model of cancers from microarray data was attempted using a fuzzy neural network and the Support Vector

The hybrid intelligent system

In this paper, a hybrid intelligent model, i.e., FMM-CART-RF, is developed to undertaking medical data classification problems. The hybrid intelligent system possesses two important properties, i.e., incremental learning with high performance and rule extraction with justifiable predictions. Fig. 1 shows the procedure of FMM-CART-RF. In order to allow FMM, CART, and RF to operate efficiently as a hybrid intelligent system, a number of modifications are needed, as explained in the following

Experimental setup

In this section, three publicly available data sets from the UCI machine learning data repository (Newman, Hettich, Blake, & Merz, 2007) were used for evaluation, i.e., Breast Cancer Wisconsin, Pima Indians Diabetes, and Liver Disorders. Before the experiments, all variables in the data set were first normalized between 0 and 1. They were then used by FMM-CART-RF for learning and prediction. A decision tree was produced to provide explanation for the predictions. The k-fold cross-validation was

Conclusions

In this paper, a hybrid intelligent model, i.e., FMM-CART-RF, has been proposed for undertaking medical decision support tasks. A series of empirical studies using three benchmark medical data sets from the UCI Machine Learning Repository, namely Breast Cancer Wisconsin, Pima Indians Diabetes, and Liver Disorders, has been conducted to evaluate the efficacy of the hybrid model. Different experimental configurations have been adopted in order to provide a fair performance comparison with

References (74)

  • E.I. Papageorgiou

    A new methodology for decisions in medical informatics using fuzzy cognitive maps based on fuzzy rule-extraction techniques

    Applied Soft Computing

    (2011)
  • K. Polat et al.

    Breast cancer and liver disorders classification using artificial immune recognition system (AIRS) with performance evaluation by fuzzy resource allocation mechanism

    Expert Systems with Applications

    (2007)
  • A. Quteishat et al.

    A modified fuzzy min–max neural network with rule extraction and its application to fault detection and classification

    Applied Soft Computing

    (2008)
  • V.F. Rodriguez-Galiano et al.

    An assessment of the effectiveness of a random forest classifier for land-cover classification

    ISPRS Journal of Photogrammetry and Remote Sensing

    (2012)
  • E.W. Saad et al.

    Neural network explanation using inversion

    Neural Networks

    (2007)
  • R. Stoean et al.

    Modeling medical decision making by support vector machines, explaining by rules of evolutionary algorithms with feature selection

    Expert Systems with Applications

    (2013)
  • J. Zhang et al.

    Random-forests-based network intrusion detection systems

    IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews

    (2008)
  • H. Abe et al.

    Implementing an integrated time-series data mining environment based on temporal pattern extraction methods: A case study of an interferon therapy risk mining for chronic Hepatitis

    New Frontiers in Artificial Intelligence, Lecture Notes in Computer Science

    (2006)
  • M.S.H. Aung et al.

    Comparing analytical decision support models through boolean rule extraction: A case study of ovarian tumour malignancy

    Advances in Neural Networks, Lecture Notes in Computer Science

    (2007)
  • A.T. Azar

    Fast neural network learning algorithms for medical applications

    Neural Computing and Applications

    (2013)
  • G. Camps-Vails et al.

    Therapeutic drug monitoring of kidney transplant recipients using profiled support vector machines

    IEEE Transactions on Systems, Man, and Cybernetics, Part C: Applications and Reviews

    (2007)
  • C.J. Carmona et al.

    Evolutionary fuzzy rule extraction for subgroup discovery in a psychiatric emergency department

    Soft Computing

    (2010)
  • G.A. Carpenter et al.

    Rule Extraction: from neural architecture to symbolic representation

    Connection Science

    (1995)
  • D. Cascio et al.

    Mammogram segmentation by contour searching and mass lesions classification with neural network

    IEEE Transactions on Nuclear Science

    (2006)
  • F. Chabat et al.

    Computerized decision support in medical imaging

    IEEE Engineering in Medicine and Biology Magazine

    (2000)
  • S. Coates et al.

    Tele-EEG in the UK: A report of over 1000 patients

    Journal of Telemedicine and Telecare

    (2012)
  • D. Coyle et al.

    Faster self-organizing fuzzy neural network training and a hyperparameter analysis for a brain–Computer Interface

    IEEE Transactions on Systems, Man, and Cybernetics, Part B: Cybernetics

    (2009)
  • G.P.K. Economou et al.

    A novel medical decision support system

    Computing & Control Engineering Journal

    (1996)
  • B. Efron

    Bootstrap methods: Another look at the Jackknife

    The Annals of Statistics

    (1979)
  • B. Efron et al.

    An introduction to the bootstrap

    (1993)
  • T.P. Exarchos et al.

    EEG transient event detection and classification using association rules

    IEEE Transactions on Information Technology in Biomedicine

    (2006)
  • S.J. Fakih et al.

    LEAD: A methodology for learning efficient approaches to medical diagnosis

    IEEE Transactions on Information Technology in Biomedicine

    (2006)
  • I.D. Falco

    Differential evolution for automatic rule extraction from medical databases

    Applied Soft Computing

    (2013)
  • L. Franco et al.

    Early breast cancer prognosis prediction and rule extraction using a new constructive neural network algorithm

    Computational and Ambient Intelligence, Lecture Notes in Computer Science

    (2007)
  • S. Ghosh-Dastidar et al.

    Principal component analysis-enhanced cosine radial basis function neural network for robust epilepsy and seizure detection

    IEEE Transactions on Biomedical Engineering

    (2008)
  • J. Han et al.

    Data mining: Concepts and techniques

    (2012)
  • F. Hayes-Roth et al.

    Building expert systems

    (1983)
  • Cited by (0)

    View full text