Elsevier

Expert Systems with Applications

Volume 40, Issue 12, 15 September 2013, Pages 5004-5010
Expert Systems with Applications

Expert system for predicting unstable angina based on Bayesian networks

https://doi.org/10.1016/j.eswa.2013.03.029Get rights and content

Abstract

The use of computer-based clinical decision support (CDS) tools is growing significantly in recent years. These tools help reduce waiting lists, minimise patient risks and, at the same time, optimise the cost health resources. In this paper, we present a CDS application that predicts the probability of having unstable angina based on clinical data. Due to the characteristics of the variables (mostly binary) a Bayesian network model was chosen to support the system. Bayesian-network model was constructed using a population of 1164 patients, and subsequently was validated with a population of 103 patients. The validation results, with a negative predictive value (NPV) of 91%, demonstrate its applicability to help clinicians. The final model was implemented as a web application that is currently been validated by clinician specialists.

Highlights

  • We have developed a Clinical Decision Support System (CDSS) with predicts unstable angina on incoming emergency patients with unspecific chest pain that are on risk of a heart attack.

  • The system achieves a 91.67% negative predictive value (NPV) over the validation dataset.

  • The CDSS is based on a Bayesian Network with 17 patient-related inputs.

  • The system runs as a web application, so that it can be used quickly and efficiently from any computer running a web browser.

Introduction

The use of computer-based clinical decision support (CDS) tools is growing in recent years due to different reasons (Steyerberg, 2009):

  • Helps the clinicians making decisions, thus reducing the clinical errors.

  • Improves the time to get a diagnostic reducing waiting time.

  • Optimizes health resources reducing unnecessary medical tests.

Those advantages have expanded the use of these tools in clinical practice in the form of web services, desktop programs or applications for mobile phones and tablets. The use of these tools in different clinical areas, supported by the increasing amount of patient data and the ability to analyse and process it through Big Data techniques, is foreseen as a huge boost in health care (Manyika et al., 2011). Currently, one of the areas that demands more resources is cardiology. Nowadays cardiovascular diseases are the main cause of death in the developed world (Escaned et al., 2008). Modern lifestyle that leads us to many stressful situations, poor diet and little exercise have made heart disease the cause of many deaths. One of the symptoms of possible heart failure angina is characterised by severe chest pain. When suffering from that pain for more than 15 min, it is highly recommended requiring clinical attention, as it can be the initial phase of a myocardial infarction. This pain occurs when the demand for oxygen by the heart muscle is not served (because there is an interruption of the blood supply to a part of the heart muscle). This pain is in fact one of the most frequent causes of admission in the Emergency Services of the hospitals. Some of these patients suffer an acute coronary syndrome that is diagnosed by electrocardiogram (ECG) findings or alteration of biomarkers of myocardial damage (troponin). However, in some patients the ECG is nonspecific and troponin is normal. This population suffers form a chest pain of uncertain origin. They are mostly low-risk patients without any heart disease, but we can not reject an acute coronary syndrome in some of them. The key challenge is to identify those patients at risk for suffering an acute coronary syndrome with normal troponin (unstable angina), within a population that is generally at low risk. The tools currently available in the emergency room of a hospital do not work, because the ECG is nonspecific, and troponin is normal. As a result, the final decision of admission or release is postponed until a treadmill stress test is carried out (usually the next morning) Sanchis et al. (2006). This strategy is suboptimal because the patient must wait for several hours, many patients can not run on the treadmill, and sometimes the results are inconclusive.

This article proposes the development of a Clinical Decision Support System, for being use in the emergency units of a hospital in order to determine the probability of unstable angina within 24 h of patient entry into the hospital. The inputs of the system are the clinical data that are routinely collected in the emergency room of a hospital. Almost all of those data have a binary response (e.g. – patient gender, smoker, etc.). With this kind of inputs, the most appropriate machine learning models are decision trees and Bayesian networks (Alpaydin, 2009). The results obtained from decision trees were not good enough (deep trees were needed, and therefore generalisation was bad). Moreover, a system whose parameters could be updated continuously when new information was available was required, that is why Bayesian networks were used. The advantages of these models for using them in clinical problems are Lucas et al. (2004):

  • The Bayesian model can be interpreted by the clinician, as the relationships among variables are clearly represented by a graph (directed graph in our model).

  • The clinician can provide expertise knowledge to establish new relationships between variables that might not have been reflected in the case of using an automatic learning algorithm for the Bayesian network structure.

  • Adding new knowledge is a straight forward process, which can be automated by updating the frequency tables of each input variable.

  • It is not necessary to know the values of all inputs to the model to obtain a valid output. Thus, if the inputs are obtained in a sequential way, as it happens in an emergency unit, or some information about the patient is missing, the system may be updating the probability as soon as new information of clinical tests is available.

Once the Bayesian network was validated, it was implemented into a Decision Support System which allowed clinicians to access it remotely, in an easy and simple way. The obvious solution was to implement the expert system as a web application. This way it is possible to centralise the data at a single location while access can be granted from any computer without requiring dedicated software (only a web browser and an Internet connection). Moreover, it is possible to implement a user-based management to control the people that use the tool and access the data. The rest of this paper is organised as follows. Section 2 explains the Bayesian models used. Section 3 discusses the data used and results obtained. Section 4 explains the Web tool development. Finally, Section 5 summarizes the conclusions of the present work.

Section snippets

Bayesian networks

A Bayesian network (BN) is a probabilistic graphical model composed of two different parts: on one hand is the graphical structure (directed acyclic graph) that defines the relationship between variables and, on the other hand, the probabilities established between these variables (Koller and Friedman, 2009, Korb and Nicholson, 2011). The elements of a Bayesian network are as follows Rusell and Norvig (2009):

  • A set of variables (continuous or discrete) forming the network nodes.

  • A set of directed

Data used, methodology and results

The implemented Clinical Decision Support System is based on a Bayesian Network model trained with a dataset of 1164 cases.

Web application

The authors have deployed a Clinical Decision Support System based on Bayesian networks as a web application. This tool evaluates the risk of heart attack on incoming patients with chest pain in an emergency unit, based on the evidences from the Patient Symptoms and his clinical history. The system consists of a web front-end with an input form for introducing and evaluating new clinical cases, a probabilistic inference motor based on a Bayesian Network (BN), and a web administration panel for

Conclusions

This paper presents a Clinical Decision Support System (CDSS) that helps the clinicians in the evaluation of incoming emergency patients with unspecific chest pain that are on risk of a heart attack. A correct diagnosis of these patients, which is difficult to achieve with the standard evaluation procedure (ECG and troponin measurements), could reduce the number of unnecessary hospitalisations. The CDSS is based on a Bayesian Network (BN) model that takes into account 17 different

References (12)

There are more references available in the full text version of this article.

Cited by (36)

  • A comprehensive scoping review of Bayesian networks in healthcare: Past, present and future

    2021, Artificial Intelligence in Medicine
    Citation Excerpt :

    In 12 papers (10%) an interface was proposed. Examples can be found in [75,57,95,48,133,126]. From the papers, where an interface was presented, only one described evaluation of the proposed interface [95].

  • Bayesian networks in healthcare: What is preventing their adoption?

    2021, Artificial Intelligence in Medicine
    Citation Excerpt :

    Evaluating the system’s impact on clinical decision making, its usefulness in practice, and the degree of practitioner acceptance were considered important and necessary steps to aid adoption in practice [80,69,62,81]. Further, BN-based systems should adhere with and be integrated into current clinical practice in order to be accepted by clinicians [72,82,47]. Even if PO barriers and facilitating factors are largely ignored, many benefits were discussed in the literature.

  • Improving Bayesian inference efficiency for sensory anomaly detection and recovery in mobile robots

    2021, Expert Systems with Applications
    Citation Excerpt :

    Finally, Section 8 summarizes the main contributions of the paper and proposes future work. Developed in the 1980s, Bayesian networks have been applied to a broad variety of fields and heterogeneous domains such as medicine (Vila-Frances et al., 2013), agriculture (Drury, Valverde-Rebaza, Moura, & de Andrade Lopes, 2017), economics (Azar & Mostafaee Dolatabad, 2019) and engineering (Sousa, Prieto-Castrillo, Matos, Branco, & Lourenço, 2018). In robotics, Bayesian network inference has been applied to a wide range of tasks, such as localization and mapping (Zhou & Sakane, 2008), navigation (Min, 2005), manipulation (Song, Ek, Huebner, & Kragic, 2015), sensory fusion (Besada-Portas, López-Orozeo, & De La Cruz, 2002), logic programming (Bessiere, 2008), etc.

View all citing articles on Scopus

This work was supported by the Spanish Ministry of Education and Science under Grant Instituto Carlos III (FEDER), Red HERACLES 06/0009.

View full text