A comparative data analytic approach to construct a risk trade-off for cardiac patients’ re-admissions

Publication Date04 Feb 2019
Pages189-209
DOIhttps://doi.org/10.1108/IMDS-12-2017-0579
AuthorMurtaza Nasir,Carole South-Winter,Srini Ragothaman,Ali Dag
subjectMatterInformation & knowledge management,Information systems,Data management systems,Knowledge management,Knowledge sharing,Management science & operations,Supply chain management,Supply chain information systems,Logistics,Quality management/systems
A comparative data analytic
approach to construct
a risk trade-off for cardiac
patientsre-admissions
Murtaza Nasir
Department of Decision Science, School of Business, University of South Dakota,
Vermillion, South Dakota, USA
Carole South-Winter
Department of Health Services Administration, School of Business,
University of South Dakota, Vermillion, South Dakota, USA
Srini Ragothaman
Department of Accounting & Finance, School of Business,
University of South Dakota, Vermillion, South Dakota, USA, and
Ali Dag
Department of Decision Science, School of Business, University of South Dakota,
Vermillion, South Dakota, USA
Abstract
Purpose The purpose of this paper is to formulate a framework to construct a patient-specific risk score
and therefore to classify these patients into various risk groups that can be used as a decision support
mechanism by the medical decision makers to augment their decision-making process, allowing them to
optimally use the limited resources available.
Design/methodology/approach A conventional statistical model (logistic regression) and two machine
learning-based (i.e. artificial neural networks (ANNs) and support vector machines) data mining models were
employed by also using five-fold cross-validation in the classification phase. In order to overcome the data
imbalance problem, random undersampling technique was utilized. After constructing the patient-specific
risk score, k-means clustering algorithm was employed to group these patients into risk groups.
Findings Results showed that the ANN model achieved the best results with an area under the curve score
of 0.867, while the sensitivity and specificity were 0.715 and 0.892, respectively. Also, the construction of
patient-specific risk scores offer useful insights to the medical experts, by helping them find a trade-off
between risks, costs and resources.
Originality/value The study contributes to the existing body of knowledge by constructing a framework
that can be utilized to determine the risk level of the targeted patient, by employing data mining-based
predictive approach.
Keywords Clustering, Decision support systems, Healthcare management,Data mining, Business analytics
Paper type Research paper
1. Introduction
1.1 Motivation
Hospital readmissions are common, costly and have received significant attention as a
correctable source of poor quality of care and excessive spending. Readmissions have an
annual cost of $19bn; a whopping 87 percent or $17bn is directly associated with additional
medical cost including ancillary services, prescription drug services, and inpatient and
outpatient care. Additional costs of $1.4bn are attributed to increased mortality rates and
$1.1bn or 10m days is associated with lost productivity from missed work based on
short-term disability claims (Andel et al., 2012). Several interventions that involve multiple
Industrial Management & Data
Systems
Vol. 119 No. 1, 2019
pp. 189-209
© Emerald PublishingLimited
0263-5577
DOI 10.1108/IMDS-12-2017-0579
Received 16 December 2017
Revised 7 March 2018
9 May 2018
Accepted 21 May 2018
The current issue and full text archive of this journal is available on Emerald Insight at:
www.emeraldinsight.com/0263-5577.htm
189
Cardiac
patients
re-admissions
components (e.g. a patient needs assessment, medication reconciliation, patient education,
arranging timely outpatient appointments and providing telephone follow-up) have
successfully reduced readmission rates for discharged patients.
Predicting patients likely to be readmitted plays a crucial role in that it enables a more
focused and cost-effective approach to decrease readmissions. Such predictive methods are
necessary to capitalize on the ubiquitous collection and storage of large amounts of data.
Concurrently, computational power has increased to reach a level where complex data
mining models (i.e. artificial neural networks (ANNs), support vector machines (SVM), etc.)
can now operate on vast amounts of data, turning data into information that can be used to
gain useful and actionable knowledge. Data mining methods have been effectively and
extensively utilized in variety of fields (Sharda and Delen, 2006; Delen, 2009; Dag et al., 2016,
2017, Berry and Linoff, 1997; Liao et al., 2012; Kankanhalli et al., 2016; Lee et al., 2017;
Oztekin and Oztekin, 2016; Topuz et al., 2017), since they enable us to characterize non-linear
dependencies between variables at an abstract and conceptual level (Sumathi and
Sivanandam, 2006). These models can also produce qualitative descriptions of regularities
and determine dependencies on factors that are not explicitly provided in the data
(Sumathi and Sivanandam, 2006). On the other hand, statistical techniques have strong
parametric assumptions about data, they have to specify a model in advance and they are
primarily oriented toward extracting statistical data characteristics, which, in turn, brings
certain inherent limitations (Sumathi and Sivanandam, 2006; Kecman, 2001). On the other
hand, machine learning models can automatically detect complex linear/non-linear relations,
by using nonparametric models (Kecman, 2001).
This study aims to develop a data mining-based methodology that will enable us to:
efficiently predict the readmission for cardiac patients upon completion of medical
procedures; develop a patient-specific probabilistic risk score to enable decision makers to
determine a trade-off between costs and benefits from a patient-care point of view; and
group the patients into risk levels based on the risk scores obtained. The overarching goal
here is to provide medical experts with a set of information, which they can effectively
utilize in their decision-making processes. In order to reach our goal, this study utilizes two
(machine learning-based) data mining techniques (i.e. ANN and SVM) and a statistical
method (Logistic Regression (LR)) to predict the readmission likelihood of cardiac patients
at the time of discharge, and to develop an associated framework that allows hospitals and
decision makers to view and use the predictions with a trade-off between available resources
and the risk. A k-means clustering application is used to stratify the patient risk predictions
to provide the decision makers an additional metric for risk assessment. Thus, the combined
framework of a machine learning prediction model combined with a clustering algorithm
applied to a healthcare decision support problem and creating a novel readmission risk
stratification methodology is the original contribution of our work. Therefore, the real
contribution of the current study is not the classification methods that have been employed,
but it is rather the end-product provided to the medical decision makers. Given the
probabilistic score that is obtained, the medical experts will be able to confidently rely on the
score provided, while there will be a small percentage of the outcomes that the medical
experts should use their own intuition/incentive to make the final decision. The remainder of
the study is as follows: in Section 1.2, the importance of the research problem as well as the
existing studies that have addressed similar problems has been discussed in detail. In this
section, both the methodological and the conceptual contributions of the current study have
been discussed clearly in order to emphasize the research gap that the current study fills.
More details about the models and the methodology that was adapted are provided in
Section 2. Section 3 presents and discusses the results that were obtained. In Section 4, the
takeaways of the current proposed methodology, and potential future research and
limitations have been briefly discussed.
190
IMDS
119,1

To continue reading

Request your trial

VLEX uses login cookies to provide you with a better browsing experience. If you click on 'Accept' or continue browsing this site we consider that you accept our cookie policy. ACCEPT