Data Mining Models to Predict Patient’s Readmission in Intensive

Care Units

Pedro Braga

, Filipe Portela

, Manuel Filipe Santos

and Fernando Rua

Information System Department, University of Minho, Guimarães, Portugal

Algoritmi Research Centre, University of Minho, Guimarães, Portugal,

Serviço de Cuidados Intensivos, Centro Hospitalar do Porto, Porto, Portugal

Keywords: Readmission, Intensive Care, INTcare, Decision Support in Intensive Care Medicine, Data Mining, SWIFT.

Abstract: Decision making is one of the most critical activities in Intensive Care Units (ICU). Moreover, it is

extremely difficult for health professionals to interpret in real time all the available data. In order to improve

the decision process, classification models have been developed to predict patient’s readmission in ICU.

Knowing the probability of readmission in advance will allow for a more efficient planning of discharge.

Consequently, the use of these models results in a lower rates of readmission and a cost reduction, usually

associated with premature discharges and unplanned readmissions. In this work was followed a numerical

index, called Stability and Workload Index for Transfer (SWIFT). The data used to induce the classification

models are from ICU of Centro Hospitalar do Porto, Portugal. The results obtained so far, in terms of

accuracy, were very satisfactory (98.91%). Those results were achieved through the use of Naïve Bayes

technique. The models will allow health professionals to have a better perception on patient’s future

condition in the moment of the hospital discharge. Therefore it will be possible to know the probability of a

patient being readmitted into the ICU.

1 INTRODUCTION

Ensuring that in a given situation are taken the better

decisions, remains a huge challenge for science and

technology. Thus, the Information Technology (IT)

has been playing a crucial role in the adoption of

measures and optimization of processes in

organizations. The field of medicine is no exception,

and the needs for rapid development and the use of

IT as a response to various problems is increasing.

As an example, there is intensive medicine, where

critically ill patients need special care for their

survival. The readmission of patients in Intensive

Care Units (ICU) is currently one of the major

problems and challenges of health professionals in

the field of intensive care medicine, because it can

be danger to the patient and spend extra resources.

Since the ICU readmissions are common and costly,

there is a need to investigate the application of new

techniques to reduce unplanned readmission of

patients into the ICU, because the currently existing

techniques do not provide satisfactory results

(Gortzis, Sakellaropoulos, Ilias, Stamoulis, and

Dimopoulou, 2008).

This study aims to demonstrate that by using data

mining (DM) techniques it was possible to

determine the probability of patient’s readmission in

order to assist health professionals in their decision

process.

This article is structured in the following way:

Introduction which is presented a brief introduction

about the environment and the foundations of the

work to develop. Background: where is presented

the problem and some theoretical fundaments:

Intensive Medicine, Readmissions, INTCare and the

Stability and Index for Load Transfer. Study

Description: describes the process of project

development. Discussion of Results: are discussed

and interpret the results obtained. Conclusions:

where are presented the most important conclusions

of the work. Future Work: presents

recommendations for future work.

604

Braga P., Portela F., Filipe Santos M. and Rua F..

Data Mining Models to Predict Patient’s Readmission in Intensive Care Units.

DOI: 10.5220/0004907806040610

In Proceedings of the 6th International Conference on Agents and Artiﬁcial Intelligence (ICAART-2014), pages 604-610

ISBN: 978-989-758-015-4

 2014 SCITEPRESS (Science and Technology Publications, Lda.)

2 BACKGROUND

2.1 Intensive Medicine

Intensive Medicine (IM) is a medical specialty

whose main goals are to diagnose and treat patients

with serious illnesses and restore them to their

previous state of health. IM can still be set up as a

“Multidisciplinary field of medical science that

specifically addresses three stages: prevention,

diagnosis and treatment of patients with potentially

reversible pathophysiological conditions that

threaten or present failure of one or more vital

functions” (Silva, 2007). Associated with IM comes

the Intensive Care Units (ICU). ICUs are

characterized as qualified locals to assume full

responsibility for patients with organ dysfunction,

supporting, preventing and reversing failure of vital

organs (Ministério da Saúde, 2003). Intensivist is a

health professional with critical care training that

works in the ICU.

2.2 ICU Readmission

An unplanned readmission of patients is directly

related to a bad decision by the intensivist at the

time of patient assessment (discharge). However, the

ability to predict relapse of a patient after the

discharge from the ICU is limited (Gajic, et al.,

2008). In order to understand how it is processed the

readmission of a patient it is important first to realize

how it is processed an admission. The admission

into UCI is, by definition, "a time of transition for

some patients whose life is at risk and it is part of a

process and not an end in itself" (Ministério da

Saúde, 2003). It is considered admission when the

patient admitted to the health facility occupies a bed

or couch for a minimum of 24 hours (ACSS, 2012).

A patient is considered readmitted if he/she is

hospitalized at the same hospital with the same

principal diagnosis within thirty days after discharge

(ACSS, 2012). According to literature review, in

North America and Europe, the average rate of

readmission of patients in ICUs is around 7%. A

study conducted by the Royal Melbourne Hospital in

Australia showed that the rate of readmission of

patients was 10.5%. The main factors can be

respiratory and cardiac problems, the progression of

the patient's condition, care needs post-operative,

and inadequate follow-up care (Russell, 1999). A

study conducted in England by SSentif Intelligence

(Intelligence, 2013), showed that on average 16% of

patients above 75 years of age suffer readmission 28

days after discharge, although this figure varies

significantly across the country, in the West South

England has an average of 12.98% and the city of

London register a value of 17.06%

2.3 Stability and Workload Index for

Transfer

It is extremely difficult for the health professionals

to interpret almost instantly all the data available. In

fact, at the time of admission or discharge of the

patient the criteria employed by the health

professional are often subjective and are not likely to

be reproduced in other cases. Many of these

professionals are often forced to rely on their

intuition and subjective analysis to assess the clinical

status of the patient and thus determine whether the

patient is ready for discharge or not (Gajic, et al.,

2008).

Published data shows that there are models or

mathematical techniques that help predict

readmission of patients in the ICU. As an example,

according to Gajic (2008), there is a study to

develop and validate a numerical index called

Stability and Workload Index for Transfer (SWIFT)

(Gajic, et al., 2008). The considered variables to be

used in SWIFT in order to estimate the probability

of unplanned readmission were: length of stay in the

ICU (duration in days), the source of the patient's

admission, the Glasgow coma scale (GCS), the

partial pressure of oxygen in arterial blood [PaO2] /

and the fraction of inspired oxygen [FIO2] and

evaluation of nursing care for respiratory problems

[PCO2].

The final score is derived from a set of

information available at the time of hospital

discharge estimating the probability of the patient in

the ICU using as support the scores presented in

Table 1.

SWIFT is according to some experts from ICU

of Centro Hospitalar do Porto (CHP) the most

popular readmission technique currently used in

Portuguese hospitals. Therefore this predictive

model was the basis of the current study using DM

techniques.

2.4 INTCare

This study is being developed under the research

project called INTCare. INTCare is an Intelligent

Decision Support System (IDSS) for Intensive Care

Medicine, and is implemented in ICU of the

Hospital de Santo António, CHP. The main

objective was to change the responsiveness of

reactive response to proactive, thus enabling

DataMiningModelstoPredictPatient'sReadmissioninIntensiveCareUnits

605

Table 1: SWIFT Variables and scores.

Variables Score

Original source of this ICU admission

Emergency department. 0

Transfer from a ward or outside hospital 8

Total ICU length of stay (duration in days)

<2 0

2-10 1

>10 14

Last measured PaO2/FIO2 ratio

>400 0

<400 and ≥150 5

<150 and ≥100 10

<100 13

Glasgow Coma Scale at time of ICU discharge

>14 0

11-14 6

8-10 14

<8 24

Last arterial blood gas PaCO2

<45 mm Hg 0

>45 mm Hg 5

physicians to ensure an improvement in the

treatment of their patients (Portela, et al., 2012). The

system uses a set of variables to try to predict the

patient's condition through intelligent autonomous

agents that are responsible for the fulfilment of

objectives, which include the use of data mining to

predict organ failure and the outcome of a patient in

the next 24 hours. Currently this approach (predict

readmissions) is part of one of the goals of INTCare

II project recently approved by the Foundation for

Science and Technology. INTCare System consists

of four subsystems within which stand the system of

data acquisition, the knowledge management,

inference and the interface. For more information

about the structure of INTCare it is possible to check

on the following article (Portela, Pinto, & Santos,

2012).

3 STUDY DESCRIPTION

3.1 Methods and Tools

For this study it was used the purpose of forecasting

through a set of classification models. For the work

were used Oracle SQL Developer for data treatment,

R for data analysis and the Oracle Data Miner

(ODM) to build prediction models (Taft, et al.,

2003). Due the type of techniques available in ODM

it was used: Support Vector Machines (SVM),

Decision Trees (DT) and Naïve Bayes (NB).

3.2 Business and Data Understanding

The Business goal of this study is decreasing the

readmissions number through the use of Data

Mining. The Data Mining objective is developing

models which allow determine the probability of

patient readmission. These models act as a support

to health professionals in decision-making and

improving the patient's health services. To this study

it was used real data provided from CHP databases.

Although the data acquisition subsystem incorporate

five different data sources (Portela, Pinto, & Santos,

2012), the attributes used were collected in real time

from only two different sources: Laboratory (LAB)

and Clinical Process (CP) – Electronic Health

Record (EHR). Table 2 presents the attributes, type,

domain and source of the information available.

Table 2: Attributes overview.

Attributes Type Domain Source

episode number - LAB/CP

nprocess number - LAB/CP

nsample varchar - LAB

daten date - CP

sex varchar [M;F] CP

po2 number

[0;

+∞ [

LAB

fio2 number [0;100] LAB

Pco2 number

[0;

+∞ [

LAB

emergencyroom varchar [True;False] CP

input date date - CP

output date date - CP

validation pco2 date - LAB

validation po2 date - LAB

To this work it was evaluated a set of episodes

with admission and discharges in the ICU, since

December 20, 2010 to April 23, 2013. The attribute

Glasgow coma scale (GCS) was not considered in

this study because it is an attribute registered

manually and consequently, countless null values

were found.

3.3 Data Preparation

At this stage it was conducted an Extract

Transforming and Loading (ETL) process, the

number of times needed until they present quality. In

this phase one of the main goal is to build new data

from the existing data. The following derivations of

new attributes were made in the dataset and can be

seen in Figure 1.

The scores come from assigning values

according to the score table SWIFT. Observing

Table 1, and ignoring the GCS, variables in the

analysis are: emergency room, length of stay,

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

606

Figure 1: Schematic construction of new data.

PO2/FIO2 and PCO2. Table 1 made a point

between variables and their respective scores

according to SWIFT. Table 1 also presented the

scores domain.

New attributes:

 Age: Domain [15 ; +∞ [

 Readmission: Domain [0,1]

 Length of Stay: Domain [1,76]

 PO2/FIO2: Domain [0, +∞ [

Scores:

 Emergency Room

 Length of Stay

 PO2/FIO2

 PCO2

In order to induce classification models some of

the input attributes needed to be transformed. The

point is to transform the attributes into discrete and

normalized values in order to be used.

For the attributes PO2/FIO2, PCO2 and Length

of Stay was necessary to convert the continuous

numeric data into class intervals. For this purpose it

was used an approach of equitable distribution of

data classes. Attributes PO2/FIO2 and PCO2 were

equitably distributed however attribute Length of

Stay had a slightly different approach.

The distribution followed the evaluation criteria

of health professionals from CHP, and the attribute

Length of Stay was divided into 3 parts: patients up

to 1 day in hospital, in other words, sick passengers

in ICU, patients between 2 and 5 days of

hospitalization, so patients without extra

complications, like infections or allergies, and

patients with more than 5 days of admission, i.e.,

patients in serious condition with high probability of

extra complications. In Table 3 it is possible to

observe the division of classes.

Table 3: Data overview of the classes created by attribute.

Attribute

Class %

PO2/FIO2

[28-120[ 14.7

[120-167[ 14.0

[167-214[ 14.1

[214-265[ 14.3

[265-328[ 14.2

[328-430[ 14.2

[430-39250[ 14.5

PCO2

[11,3-32,5[ 14.0

[32,5-35,9[ 14.2

[35,9-38,8[ 13.9

[38,8-41,4[ 14.0

[41,4-44,7[ 14.3

[44,7-50[ 14.1

[50-134] 15.5

Length of

Stay

[1] 34.4

[2-5[ 37.8

[5-76] 27.8

3.4 Modelling

This phase focused on getting models to translate

business goals through the application of data

mining techniques. The process suffered two

approaches, one for the normal table generated

(score_swift_A) and one for a second table

(score_swift_B) which suffered an oversampling.

The necessity of using the oversampling technique is

related to the existent dominance of a type of data

distribution over the other. When facing this

imbalance, the models tend to have a greater number

of errors over the minority of the data distribution.

So the oversampling through the data replication,

increase the weight of minority data distribution, in

order to be able to obtain a better performance of the

classification models (Barua, Islam, and Murase,

2013).

In order to understand what has been done in the

approach (A) the balance is 98.66% for no

readmitted patients and 1,34% for readmitted

patients. For the 2

approach (B) the balance is 51%

for no readmitted patients and 49% for readmitted

patients. This takes into account two different

approaches and eight scenarios using or not

CaseMix.

CaseMix is generally represented as non-clinical

data collected from the patient which may be useful

for analysis and diagnosis or for research purposes.

For the design and evaluation of the models were

used 70% of the data for training and the remaining

DataMiningModelstoPredictPatient'sReadmissioninIntensiveCareUnits

607

30% for the test. The training block is used to

produce the model and the test is used to measure

the ability of generalization of the model. The test

block is also the percentage of data that are not used

in the learning process. The encoding model is

presented as follows:

Target (dependent) =

eadmission;

Input (independent) = {Normal, Scores, Class

and CaseMix};

Normal =

{pco2, po2/fio2, emergencyroom,

len

th of sta

Scores =

{pco2_score, po2/fio2_score,

emergencyroom_score, lengthof

sta

score

}

Classes =

{class_pco2, class_po2/fio2,

class

len

thofsta

, emer

enc

room

}

CaseMix = {sex, age}

Models applied to the two approaches are

presented as follows:

M1 = {Normal + Scores}

M2 = {Scores}

M3 = {Normal}

M4 = {Normal + Scores + Class}

M5 = {Normal + Scores + CaseMix}

M6 = {Scores + CaseMix}

M7 = {Normal + CaseMix}

M8 = {Normal + Scores + Class + CaseMix}

Taking into account the two approaches it was

developed a total of 48 models: 2 approaches * 8

scenarios * 3 techniques * 1 target.

In the construction of classification models the

parameters were catalogued using oracle data

mining according to the representation in Table 4.

Table 4: Techniques Settings.

Description Values

Algorithm Name Support Vector Machine

Kernel Function Gaussian

Tolerance Value 0,001

Active Learning yes

Algorithm Name Decision Tree

Homogeneity Metric Gini

Maximum Depth 7

Algorithm Name Naive Bayes

Pairwise Threshold 0

Singleton Threshold 0

3.5 Evaluation

The evaluation phase focuses primarily on the

assessment and comparison of the results, noting

whether they meet the initial goals of the project. All

models were reviewed and evaluated in order to find

influencing factors which may have escaped at

earlier stages. In order to prove the results for the set

of models developed for the readmission of patients,

it is possible to observe in Tables 5 and 6 the

attained results in terms of the three metrics:

accuracy, sensitivity and specificity. In Table 5

there are the best three models in terms of accuracy.

As it is possible observe the approach B is the one

that presents the best results and the Naive Bayes

figure as the technique that reaches the highest

values of acuity most precisely with models

NB_M7B and NB_M5B. Also in line with the

results obtained, it is possible to observe through

Table 6 the best 3 models for each metric and

approach.

Table 5: Best models results.

Accuracy Sensitivity Specificity

1º NB_M7B 98.91% 99.67% 98.12%

2º NB_M5B 98.12% 98.02% 98.23%

3º DT_M5B 96.77% 98.51% 95.02%

Table 6: Best models for approach A and approach B.

A Sensitivity Specificity Accuracy

1º

DT_M4A

99.63%

SVM_M4A

71.56%

NB_M5A

94.07%

2º

DT_M1A

99.52%

SVM_M2A

47.48%

NB_M7A

92.51%

3º

SVM_M8A

99.52%

SVM_M1A

42.92%

NB_M8A

91.96%

B Sensitivity Specificity Accuracy

1º

NB_M7B

99.67%

NB_M5B

98.23%

NB_M7B

98.91%

2º

DT_M5B

98.51%

NB_M7B

98.12%

NB_M5B

98.12%

3º

DT_M7B

98.51%

SVM_M7B

95.77%

DT_M5B

96.77%

In order to evaluate the results and to compare

the classification models for patients at the ICU

readmission, an analysis was performed through the

Receiver Operating Characteristic (ROC). The

analysis via ROC curves, which were recently

introduced in the field of DM (Prati, Batista, &

Monard, 2008) have become a powerful tool in the

evaluation of classification models. Currently, the

ROC curves are widely used in medicine, taking into

account its extreme importance in the evaluation of

medical treatments. A point near 100% is the perfect

model, i.e., all positive and negative examples are

correctly classified. From Figure 2 it was observed

that the model NB_M7_CASEMIX_B is what is

closest to the perfect spot, presenting itself as the

best classification model.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

608

Figure 2: ROC for model

NB_M7_CASEMIX_B.

4 DISCUSSION OF RESULTS

The best classification model in terms of accuracy is

the NB_M7B with a precision of 98.91%, followed

by NB_M5B with 98.12% and DT_M5B with 96,

77%.

The results showed in table 6 corroborated the

idea of being necessary the use of oversampling

techniques. As can be observed for example to the

models without oversampling the best specificity

result it is 71.56% and for the models with

oversampling it is 98.23%. In fact before the use of

the oversampling technique (A) the results were

poor, however after the implementation of

oversampling (B) the results improved significantly.

One thing that is noticeable is that most of the top

ranking models in this study are given through the

technique of DM Naive Bayes. Based on the results

obtained so far it is possible conclude that to obtain

good results it was important to include other non-

clinical attributes of the patient like age and sex.

5 CONCLUSIONS

This work contributed with a set of classification

models that can help health professionals in the

decision making. The models were introduced in

order to predict whether a patient will be readmitted,

or whether a patient won’t be readmitted. The

adoption of these models may result not only in

reducing the high mortality associated to early bad

discharges, but also can contribute to a reduction of

costs of institutions. Furthermore, the impact of

these models on the ICU will be huge, because

allows the health professionals to use them as a

decision support and, consequently, improve the

health care of patients.

6 FUTURE WORK

Completed this study it is important to draw some

prospective directions. Future work will include: To

reate new models through the exploration and

inclusion of new

attributes; To determine an

evaluation metric of the models; To create patterns

(clustering) for readmission through the generated

models and integrate the developed models in a

decision support system, in particular INTCare and

optimize the developed models.

ACKNOWLEDGEMENTS

This work is supported by FEDER through

Operational Program for Competitiveness Factors –

COMPETE and by national funds though FCT –

Fundação para a Ciência e Tecnologia in the scope

of the project: FCOMP-01-0124-FEDER-022674.

The authors would like to thank FCT

(Foundation of Science and Technology, Portugal)

for the financial support through the contract

PTDC/EIA/72819/ 2006 (INTCare) and PTDC/EEI-

SII/1302/2012 (INTCare II).

REFERENCES

ACSS, 2012. Administração Central do Sistema de Saúde,

Circular Normativa nº 33/2012 Available at:

http://www.acss.min-saude.pt.

Barua, S., Islam, M. & Murase, K., 2013. ProWSyn:

Proximity Weighted Synthetic Oversampling

Technique for Imbalanced Data Set Learning.

Advances in Knowledge Discovery and Data Mining.

s.l.:Springer Berlin Heidelberg, pp. 317-328.

Chapman, P. et al., 1999. CRISP-DM 1.0, s.l.: SPSS.

Gajic, O. et al., 2008. The Stability and Workload Index

for Transfer score predicts unplanned intensive care

unit patient readmission: Initial development and

validation. Crit Care Med. 676-682.

Gortzis, L. G. et al., 2008. Predicting ICU survival: A

meta-level approach, s.l.: BMC Health Services

Research.

Intelligence, S., 2013. Hospital Readmissions on the

Increase, press release.

Ministério da Saúde, 2003. Cuidados Intensivos:

Recomendações para o seu desenvolvimento.

Portela, F., Pinto, F. & Santos, M. F., 2012. Data Mining

Predictive Models for Pervasive Intelligent Decision

Support in Intensive Care. Advances in Information

Systems and Technologies.

Portela, et al., 2012. Intelligent and Real Time Data

Acquisition and Evaluation to Determine Critical

Events in Intensive Medicine. Procedia Technology.

DataMiningModelstoPredictPatient'sReadmissioninIntensiveCareUnits

609

Prati, Batista & Monard, 2008. Curvas de ROC para

avaliação de classificadores. pp. 215-222.

Russell, S., 1999. Reducing readmissions to the intensive

care unit.. Heart Lung.

Santos, M. F., Mathew, W. & Portela, C. F., 2011. Grid

data mining for outcome prediction in intensive care

medicine, s.l.: Springer CCIS.

Silva, A., 2007. Modelos de Inteligência Artificial na

análise da monitorização de eventos clínicos adversos,

Disfunção/Falência de órgãos e prognóstico do

doente, Universidade do Porto, PHD Thesis;

Taft, M. et al., 2003. Oracle Data Mining Concepts, 10g

Release 1 (10.1), s.l.: Oracle Corporation.

ICAART2014-InternationalConferenceonAgentsandArtificialIntelligence

610