Real-Time Data Mining Models for Predicting Length of

Stay in Intensive Care Units

Rui Veloso

, Filipe Portela

, Manuel Filipe Santos

, Álvaro Silva

Fernando Rua

, António Abelha

and José Machado

Algoritmi Centre, University of Minho, Guimarães, Portugal

Serviço Cuidados Intensivos, Centro Hospitalar do Porto, Hospital Santo António, Porto, Portugal

CCTC, University of Minho, Braga, Portugal

Keywords: Length of Stay, INTCare, Intensive Care Units, Data Mining, Real-Time.

Abstract: Nowadays the efficiency of costs and resources planning in hospitals embody a critical role in the

management of these units. Length Of Stay (LOS) is a good metric when the goal is to decrease costs and to

optimize resources. In Intensive Care Units (ICU) optimization assumes even a greater importance derived

from the high costs associated to inpatients. This study presents two data mining approaches to predict LOS

in an ICU. The first approach considered the admission variables and some other physiologic variables

collected during the first 24 hours of inpatient. The second approach considered admission data and

supplementary clinical data of the patient (vital signs and laboratory results) collected in real-time. The

results achieved in the first approach are very poor (accuracy of 73 %). However, when the prediction is

made using the data collected in real-time, the results are very interesting (sensitivity of 96.104%). The

models induced in second experiment are sensitive to the patient clinical situation and can predict LOS

according to the monitored variables. Models for predicting LOS at admission are not suited to the ICU

particularities. Alternatively, they should be induced in real-time, using online-learning and considering the

most recent patient condition when the model is induced.

1 INTRODUCTION

With a constant increase in health expenses and the

aggravation of the global economic situation,

managing costs and resources in healthcare assumes

an important role. In Intensive Care Units (ICU) this

reality is even more felt because of the costs

associated to each patient. These costs are quite high

due to several factors as is the large number of

equipment used, the high number of nurses and

doctors, the great variety of therapies and drugs

administered and the patient Length of Stay (LOS).

All these factors lead to an increase of daily patient

costs in an ICU. Planning and provisioning

resources and costs in ICU assumes a critical role in

hospital management (Marshall et al., 2005).

Accurate predictions of the patient's LOS in the ICU

can contribute to reduce costs. It is recognized by

the ICU professionals the importance of having a

tool which help to predict the patient's LOS. This

can help managing the resources and better

understand the patient condition, contributing to

avoid wrong discharges.

Related works found in literature are not accurate

for ICU setting, due to patient specific conditions.

Consequently, the principal motivation of this

research is to induce data mining models that can

predict accurately the LOS of ICU inpatients.

The idea of this work is to understand how many

days a patient with some clinical diseases need to be

in the unit. This only it is possible having a system

which can in real-time adapt the patient's LOS

according to the patient condition. With a correct

prediction of LOS it is possible make a better

resources allocation and consequently decrease the

patient costs, for example, the unit will have a better

professional’s allocation and the therapeutics

(dosages and schedules) can be prescribed according

to the patient LOS. To reach the objectives, two

approaches were explored using Data Mining (DM)

in order to predict the probability of a patient be

discharged in the next hours. The first approach (A)

considered data acquired in the first twenty-four

245

Veloso R., Portela F., Filipe Santos M., Silva Á., Rua F., Abelha A. and Machado J..

Real-Time Data Mining Models for Predicting Length of Stay in Intensive Care Units .

DOI: 10.5220/0005083302450254

In Proceedings of the International Conference on Knowledge Management and Information Sharing (KMIS-2014), pages 245-254

ISBN: 978-989-758-050-5

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

hours and the second approach (B) used data

acquired in real-time. Although the two approaches

have the same goal it is very difficult to compare

them, because the way how the goal was achieved

differs. The approach A predicted a LOS class and

the Approach B predicted the probability of a patient

be discharged in the next hour. The second approach

was explored only because approach A failed. For a

better understanding of the work, the two

approaches will be evaluated separately.

Approach A used data from the patient

admission plus some clinical variables and presented

weak results. These models used classification

techniques to predict the best LOS class. The overall

accuracy of these models ranged from 7.60% to

73.28%. For approach A is not possible to determine

sensitivity and specificity because involves more

than two output classes. Sensitivity has a great

importance in clinical decision because physicians

look for models sensible to one result. The models,

which predict patient discharge hourly, revealed

good results. Mostly all of the models presented

accuracies upper than 70% and in some cases

sensitivities upper than 90%. This work was based in

the study started by INTCare project.

This article is divided into five sections. The first

one introduces the paper. The main concepts and

related work are described in the background in the

section two. In the third one is described the study,

data analysis, transformation, modelling and results.

The subsequent section discusses the results

obtained. The last section makes final considerations

and proposes future work.

2 BACKGROUND

2.1 Length of Stay

Expenditures in hospital care are rising and the costs

associated to the patient represent a significant

portion of the expenses. Reducing Length of Stay

(LOS) in hospitals will lead to an effective reduction

in the costs and an effectiveness improvement in

hospitals (Clarke, 1996). The LOS is the number of

days that a patient is subjected to treatment since

their admission date until the discharge date. This

number of days is counted if the patient is at home

or at the hospital, depending where he is receiving

health cares. The use of this indicator permits to see

the costs with the patient or group of patients along

as the quality of the services given to the patients

(Marshall et al., 2005). Patients with outsized LOS

in the ICU represent an enormous usage of resources

so a reliable prediction of LOS can lead to a rational

bed allocation and to a resources maximization

(Wagener et al., 2011), being more efficiently.

Several works can be found about LOS but for

predicting LOS in ICU no one was found. Isken and

Rajagopalan (2002) use clustering techniques and k-

means algorithms to identify types of patients so that

can be constructed computational simulations or

analytic models about patient flow inside hospitals.

These flow models consider the analysis of measures

like LOS and bed occupancy rate. Another work

used classification techniques recurring to decision

three algorithms, support vector machines and

artificial neural networks to predict LOS in patients

with cardiac disease (Hachesu et al., 2013). Azari

(Azari et al. 2012) propose a multi-layer approach to

predict LOS in hospitals. The work used clustering

techniques to build the training data set. Then,

several classification algorithms were used with the

Naïve Bays, SVM, JRIP and J48 showing the best

results. In the area of Gastroenterology (Zhang et al.

2012) were developed models to predict LOS in ICU

in post-operatory elderly patients with gastric

cancer. The study used regression trees to predict

LOS and other hospital metrics. Caetano et al.

(2014) proposed a data-driven predictive model to

LOS with data from Portuguese hospitals from 2000

to 2013 and used as inputs the typical available

indicators at the hospitalization process. The work

developed several DM Models with the best results

being achieved with Random Forest methods.

2.2 INTCare

This study is being developed under the research

project INTCare. INTCare is a Pervasive Intelligent

Decision Support System (PIDSS) that is in constant

testing and development. Actually it is implemented

in the ICU of the Hospital de Santo António, Centro

Hospitalar Porto. This PIDSS is based on intelligent

agents (Santos et al., 2011) and aims to support the

decision making process and predict clinical events

anywhere and anytime like patient organ failure

(cardiovascular, respiratory, renal, hepatic,

neurological and hematologic), patient outcome

(Portela et. al, 2013a), readmissions, medical

diseases and others. Regarding the predictions made

the system is able to support the decision on

procedures, treatments and therapies. Predictions are

made by means of DM techniques. The system is

based in four autonomous subsystems (data

acquisition, knowledge management, inference and

interface) that use intelligent agents to perform their

actions (Portela et al. 2013b, Portela et al., 2012).

KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing

246

2.3 Data Mining

Data Mining (DM) is a field that uses techniques and

methods in data to find trends or patterns in order to

support the discovery of new knowledge (Santos &

Azevedo, 2005). DM identifies opportunities or

threats in business, to support decision making

process or to solve problems. Among the best-

known application areas are Customer Relationship

Management (CRM), banking, retail and logistics

businesses, the entertainment industry, sports,

insurance, travel industry (airlines, hotels, car

rental), health and medicine. A classification model

can be viewed as a function for mapping new cases

(input) to a class from a list of classes (output)

(Santos & Azevedo, 2005). For example, these type

of models can be used to classify the banking history

of users as good or bad (Rokach & Maimon, 2010).

2.3.1 Decision Trees

Decision Trees (DT) was introduced in 1986 in the

machine learning area with the use of ID3 algorithm

to synthesizing decision trees (Quiland, 1986). A DT

based classifier is a partition of an instance of a

space. The decision trees are composed by nodes

and branches. Usually, the DT is constructed with a

top-down strategy with the first node, the root node,

don’t having incoming nodes and giving origin to

two or mode nodes according to the attributes that

are being used. The remaining nodes have only one

incoming node. The nodes that give origin to others

are called internal or test nodes and the others are

called leaves or terminal nodes (Hand, 2001). Figure

1 represents the concept of DT.

Figure 1: Decision Tree, adapted from (Kantardzic, 2011).

The models developed using DT technique

produced rules in the form of IF predictive

information THEN target (Taylor, 2010). These

models use for example CART algorithm that is a

DT learning technique to produce DT as a binary

recursive partitioning procedure capable of

processing continuous and nominal attributes as

targets and predictors (Breiman et al, 1984).

2.3.2 Naïve Bayes

One of the first implementations of a Naïve Bays

classifier occurred in 1994 (Michie et al, 1994).

Bayesian classification is a technique to the

classification of patterns in a set of data. This type of

technique assumes that the classification of patterns

is expressed in probabilistic terms. The classification

based on Bayes theory pretend to classify objects

based on the statistical information of the objects to

minimize the probability of the classification is done

poorly (Cios et al. 2007). The Bayes theorem can be

expressed by the following expression:

P(Ci|X)=(P(X|Ci)*P(Ci))/P(X) (1)

Where,

P(Ci|X) represents the a posteriori probability;

P(Ci) represents the a priori probability;

P(X|Ci) probability density function ( );

P(X) unconditional probability density function.

2.3.3 Support Vector Machines

Support Vector Machines (SVM) are classification

methods to map vectors into a space using nonlinear

mapping building an optimal separating hyperplane

(Vapnik, 1995). The optimal hyperplane is obtained

when the hyperplane is as far as possible from all the

data points that compose the space. SVM algorithms

extend the measurement space in more complex

surfaces so the space includes combinations of the

raw variables. Then trough the margin, the algorithm

attempt to optimize the location of the linear

decision boundary between two classes that is likely

to lead to the best possible generalization, using the

biggest margin (Hand et al, 2001). Figure 2

represents a set of data that can be linearly separable

and where exists an infinite number of hyperplanes

and decision boundaries.

Figure 2: Hyperplanes and Decision Boundaries, adapted

from (Han, 2011).

X>1

Y=?

Class1 Class2 Class3 Class4

Yes

Y=A

Y=B

Y=C

Real-TimeDataMiningModelsforPredictingLengthofStayinIntensiveCareUnits

247

2.3.4 CRISP-DM

Once this study is a data mining work it was

followed the CRISP-DM methodology. CRISP-DM

is an acronym for Cross Industry Standard Process

for Data Mining. This paper is organized around the

first five tasks of this methodology: business

understanding, data understanding, data preparation,

modelling and evaluation (Wirth & Hipp, 2000).

3 STUDY DESCRIPTION

3.1 Business Understanding

As mentioned above, the prediction of LOS assumes

an important role on the optimization of costs and

resources in the ICU. In these units a lot of attributes

are available to be used for predicting LOS. Two

approaches were experimented in this study for

predicting LOS of patients in the ICU. The first

approach (A) uses as input data values of collected

in the first 24 hours considering data like the

admission type (worst vital signs values). This

approach is transversal to many areas in medicine

but is a bad choice when used to predict LOS in

intensive medicine. Alternatively, a second approach

(B) was explored. This approach considers more

variables and intends to predict the LOS based in the

probability of a patient be discharged in the next

hour. Then, the LOS is obtained by finding the hour

with the higher probability of the patient be

discharge. This hour represents the inpatient time.

Both approaches used Oracle Data Mining

(ODM) and considered three of the classification

DM techniques available: Decision Trees, Support

Vector Machines and Naïve Bayes.

3.2 Data Understanding and

Transformation

3.2.1 Approach A

The acquisition of the data used in this approach is

made from the electronic health record, laboratory,

and vital signs monitor. Table 1 presents some

statistics about the variables considered.

Case Mix consists in a set of variables that are

present in the Electronic Health Record of the

patient. In this approach were included the variables

age, admission type and admission from, admission

with surgery, Glasgow coma scale, hematologic

disease, transplanted, HTA in treatment, insulin

Table 1: Variables used in approach A.

ID Variable Min Max Value

Age Age 20 96 Real

Sex Sex M F 1 or 2

Admission Type

Urgent - - U

Programmed - - P

Admission surgery

With - - W

Without - - WO

Admission From

Chirurgic - - 1

Observation - - 2

Emergency - - 3

Nursing Room - - 4

Other ICU - - 5

Other Hospital - - 6

Other - - 7

Glasgow Coma GCS 3 15 Real

Hematologic

Disease

Yes - - Y

No - - N

Transplanted

Yes - - Y

No - - N

HTA in Treatment

Yes - - Y

No - - N

Insulin Treated

Diabetes

Yes - - Y

No - - N

Non-Insulin

Treated Diabetes

Yes - - Y

No - - N

Days of Inpatient Inpatient Days 0 40 Real

Hours of Inpatient

Inpatient

Hours

2 970 Real

Oxygen Saturation SPO2 86.00 100.00 Float

Systolic Blood

Pressure

SYS 80.64 219.48 Float

Diastolic Blood

Pressure

DIA 43.20 179.88 Float

Corporal

Temperature

TEMP 34.16 39.74 Float

Heart Rate

Measure

EGG_HR 50.00 231.00 Float

treated diabetes, non-insulin treated diabetes, hours

and inpatient days.These variables are obtained in

the moment of admission. Other variables were

considered (physiological variables) and collected

during the first 24 hours. From the vital signs

monitor, the following variables were collected:

Oxygen Saturation (SPO2), Systolic Blood Pressure

(SYS), Diastolic Blood Pressure (DIA), Corporal

Temperature (TEMP) and Heart Rate Measure

(EGG_HR). In this approach, they were introduced

some classes for the variables days and inpatient

hours (2 classes for each variable), EGG_HR, SPO2,

ART_SYS and ART_DIA by dividing the cases of

each class by equal number. The variables age, sex,

admission type, admission with surgery, admission

from, Glasgow coma scale, hematologic disease,

transplanted, HTA in treatment, insulin treated and

KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing

248

non-insulin treated were considered in the Case Mix

group.

3.2.2 Approach B

The acquisition of the data used in this approach is

made from three different sources: the electronic

nursing records, bedside monitors and laboratory. A

pre-processing agent is responsible for validating all

the collected data and for preparing data to the data

mining input table (DMIT). This agent creates the

model input structure for a patient, validates the

input sources data and processes variables.

For this approach where considered 6 main

groups of attributes:

 Case Mix (CM) – Age, admission type,

admission from;

 SOFA – Cardiovascular, Respiratory, Renal,

Liver, Coagulation and Neurologic;

 ACE (Accumulated Critical Events) – ACE of

Blood Pressure (BP), ACE of Oxygen

Saturation (SPO2), ACE of Heart Rate (HR)

and Total ACE;

 Ratios 1 (R1) – ACE of BP / elapsed time of

stay, ACE of SPO2 / elapsed time of stay,

ACE of HR / elapsed time of stay, Total of

ACE / elapsed time of stay;

 Ratios 2 (R2) – ACE of BP / max number of

ACE of BP, ACE of SPO2 / max number of

ACE of SPO2, ACE of HR / max number of

ACE of HR;

 Ratios (R) – Union of the two sets of ratios (R1

and R2).

Case Mix consists on a set of variables that are

present in the Electronic Health Record (EHR) of

the patient. In this approach it was considered the

variables age, admission type and admission from.

These variables are obtained at patient admission

and are automatically transformed according to the

DM attributes.

The Sequential Organ Failure Assessment

(SOFA) is used in ICU to score the degree of

dysfunction/failure of the cardiovascular,

respiratory, renal, liver, coagulation and

neurological organic systems. The value 0 represents

normal function and the maximum value 4 stands for

organ failure. In this case the variable used only

consider two values 0 (sofa = 0) or 1 (sofa > 0).

Accumulated Critical Events (ACE) were

considered as a complement because SOFA has no

predictive capability. ACE include four

physiological variables: blood pressure (BP),

Oxygen Saturation (SPO2) and Heart Rate (HR).

ACE are calculated by summing the critical event

values by hour.

The groups of ratios were added in sequence of

the ACE group. These ratios allow determining the

number of ACE per hour (R1) and a correspondence

between the number of ACE and the maximum

number of events verified in the past, grouped by

category and by patient (R2).

Table 2 represents the transformations on the

data for this approach as well the variables

considered. Note that for Case Mix attributes it is

only used one value to represent each case and for

the SOFA attributes is considered the worst SOFA

value received hourly.

Table 2: Variables used and transformed in approach B.

ID Variable Mi Ma Value

Age Age

18 46 1

47 65 2

66 75 3

76 130 4

Admission Type

ent - - U

Pro

ramme - - P

Admission From

Chirur

ic - - 1

Observatio - - 2

Emer

enc

- - 3

Nursin

- - 4

Other ICU - - 5

Other - - 6

Othe

- - 7

SOFA

Cardiovascula

(

mean

)

0 70 1

amine 0.0 - 1

Renal

Dobutamin 0.0 - 1

i / 0.0 - 1

Respiratory

Creatinine 1.2 - 1

Po2/Fio2 0 400 1

atic Bilirubin 1.2 - 1

Coa

ulation Platelets 0 150 1

ACE 0 +∞ SET

R1 0 1 SET

R2 0 1 SET

R1, R2 and ACE values were discretised. These

values were categorized and grouped considering an

interval of values, minimum and maximum. The sets

created were defined considering the average and the

higher value of data collected. The ranges were

created considering the Clinical Global Impression –

Severity Scale (CGI-S) (Guy, 2000). The criterion

used to define percentages concentrate the most part

of patient values within a scale between 0 and 5.

Levels 6 and 7 are for severe cases. Table 3 shows

the discretization rules.

Table 4 represents the discretization of the

variables. On top there are the different sets and on

the left the variables affected.

R1 classes are determined by the rows R1 BP

Min to R1 TOT Max and the R2 attributes (BP, O2,

HR and Total) follow the same rule, e.g. for the first

Real-TimeDataMiningModelsforPredictingLengthofStayinIntensiveCareUnits

249

Table 3: Discretization rules for approach B.

Set

R1 R2

Definition Average Maximum

> <= > <=

0 - 0% - 0%

nexistence

1 0% 25% 0% 10%

ormal Condition

2 25% 50% 10% 25%

orderline Condition

3 50% 100% 25% 50%

ild Condition

4 100% 150% 50% 75%

oderate Condition

5 150% 200% 75% 90%

arked Condition

6 200% 300% 90% 100% Severe Condition

7 300% 1000% 100% 200%

xtreme Condition

level (1) the range is from 0.00 (0%) to 0.10 (10%).

ACE attributes are grouped by their importance

and number of occurrences. These values where

defined by ICU experts but can be modified in future

studies. All these values are calculated in real-time

according to the most recent patient results. The

values presented in table 4 are the values verified in

the moment when the study was performed. These

new variables used by the models have as prefix

“C_”, e.g. for R2 the class of BP_Max is represented

by C_BP_Max.

Table 4: Discretization sets of Data Mining Input.

SET 0 1 2 3 4 5 6 7

R1 Min -0.100 0.000 0.010 0.021 0.041 0.062 0.082 0.123

BP Max 0,000 0.010 0.021 0.041 0.062 0.082 0.123 2.000

R1 Min -0.100 0.000 0.018 0.036 0.072 0.108 0.144 0.216

O2 Max 0,000 0.018 0.036 0.072 0.108 0.144 0.216 2.000

R1 Min -0.100 0.000 0.004 0.008 0.015 0.023 0.030 0.045

HR Max 0,000 0.004 0.008 0.015 0.023 0.030 0.045 2.000

R1 Min -0.100 0.000 0.020 0.041 0.081 0.122 0.162 0.243

TOT Max 0.000 0.020 0.041 0.081 0.122 0.162 0.243 2.000

Min -0.100 0.000 0.100 0.250 0.500 0.750 0.900 1.000

Max 0 0.100 0.250 0.50 0.750 0.900 1.000 2.000

ACE

Min -0.100 0 3 5 8 10 12 15

Max 0 3 5 8 10 12 15 50

Table 5: Data Analysis.

Variable Min Max AVG STDDEV VC (%)

BP 0 42 0.430 1.054 245.070

BP_Max 0 4.400 0.083 0.230 276.948

BP_Hou

0 0.500 0.008 0.025 331.631

0 16 0.500 1.474 294.800

HR_Max 0 1 0.051 0.140 275.827

HR_Hou

0 0.500 0.008 0.025 305.664

O2 0 42 1.030 2.872 278.835

O2_Max 0 1 0.034 0.091 267.398

O2_Hou

0 1 0.021 0.055 267.828

Total_ACE 0 50 1.970 4.340 220.305

Total_Ma

0 1.170 0.052 0.108 207.934

Total_Hou

0 1 0.0370 0.077 207.751

Table 5 presents the distributions of the numeric

fields used in approach B. It is possible to observe

that the sample data have quite heterogeneous

distributions because the variation coefficient is

lesser than 20%. This value is used to assess if a

distribution is hetero or homogeneous. Table 5

presents the minimum value (MIN), maximum value

(MAX), average (AVG), standard deviation

(STDDEV) and coefficient of variance (VC).

Table 6 presents the distribution of the classes

(in percentage) for each one of the independent

variables.

Table 6: Classes Distribution.

Variable Values Percenta

Admission Type

P 23.5%

U 76.5%

Age

1 16.9%

2 36.4%

3 20.8%

4 25.9%

Res

irator

1 61.6%

Coa

ulation 1 39.0%

Renal 1 19.7%

atic 1 17.6%

Cardio 1 61.6%

Admission From

1 47.3%

2 0.3%

3 17.7%

4 14.9%

5 2.2%

6 2.0%

7 15.5%

Figure 3 represents the target (discharge)

distribution. The discharged patients are represented

by the value 1, and by 0 are the ones that are

hospitalized.

Figure 3: Target Distribution.

3.3 Modelling

3.3.1 Approach A

In this approach 24 models were created combining

2 different scenarios, 4 target and 3 techniques.

KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing

250

Scenarios:

1 = {CM, Class EGG_HR, Class SPO2, Class

SYS, Class DIA, Class Temp}

2 = {CM, EGG_HR, SPO2, SYS, DIA, Temp}

Target:

1 = Class Hours 1

2 = Class Days 1

3 = Class Hours 2

4 = Class Days 2

Techniques:

1 = Support Vector Machine

2 = Decision Trees

3 = Naïve Bayes

Each one of the targets was defined according to

Sturges Rules or with clinical domain provided by

ICU physicians.

Class Hours 1 – Sturges Rules

[2, 100.67]

] 100.67, 199.39]

] 199.39, 297.98]

] 297.98, 396.65]

] 396.65, 495.32]

] 495.32, 593.99]

] 593.99, 692.66]

] 791.33, 890]

] 890, 970]

Class Hours 1 – Clinical domain

[0, 72]

] 72, 120]

] 120, +∞ [

Class Days 1 – Sturges Rules

[0, 4.4]

] 4.4, 8.8]

] 8.8, 13.2]

] 13.2, 17.6]

] 17.6, 22]

] 22, 26.4]

] 26.4, 30.8]

] 30.8, 35.2]

] 35.2, 40]

Class Days 2 – Clinical Domain

[0, 3]

] 3, 5]

] 5, +∞]

In order to evaluate the models created, 70% of

the data was considered for training and 30% for test

(hold-out sampling). The data used in models

corresponds to admissions / discharges made in ICU

of CHP from 2011.08.18 to 2014.02.08 (905 days),

to 407 patients and to 448 records. The

configurations used for these models were based in

the ODM pre-defined settings.

3.3.2 Approach B

For approach B were developed 39 models. These

39 models combine 13 different scenarios, 1 target

and 3 different techniques.

Scenarios:

1 = {CM}

2 = {CM + ACE + R}

3 = {CM + ACE + R1}

4 = {CM + ACE + SOFA}

5 = {CM + ACE + SOFA + R}

6 = {CM + ACE + SOFA + R1}

7 = {CM + ACE + SOFA + R2}

8 = {CM + ACE-SET + R-SET}

9 = {CM + ACE-SET + R1-SET}

10 = {CM + SOFA + ACE-SET}

11 = {CM + SOFA + ACE-SET + R-SET}

12 = {CM + SOFA + ACE-SET + R1-SET}

13 = {CM + SOFA + ACE-SET + R2-SET}

Target:

1 = Discharge

Techniques:

1 = Support Vector Machine

2 = Decision Trees

3 = Naïve Bayes

The models created used the same configurations

of approach A: 70% of the data for training and 30%

for testing. The data used in the models correspond

to admissions / discharges made in ICU of CHP

from 2012.02.01 to 2014.04.24 (813 days),

considering 526 patients and 55442 records. Each

one of the models has been induced automatically

and in real-time using streamed data. The data

mining engine used the data present in the input

table. This data can be represented as a tuple:

DMIT=<pid, date, hour, V_ace_bp,

V_aceBP_time, V_aceBP_max, V_ace_hr,

V_acehr_time, V_aceHR_max, V_ace_spo2,

V_ace_spo2_time, V_ace_spo2_max, V_total_ace,

V_total_ace_time, V_total_ace_max, V_age,

V_admin_F, V_admin_T, V_sofa_cardio,

V_sofa_resp, V_sofa_renal, V_sofa_coag,

V_sofa_hepa>

In this tuple pid represents the patients’

identification, the date the date of the values, hour

represents the number of hours elapsed since the

patient is admitted in the ICU and the tuples with V

prefix represent the values obtained for each patient

in each date / hour. Table 7 presents the

configurations for each one of the algorithms. For

Real-TimeDataMiningModelsforPredictingLengthofStayinIntensiveCareUnits

251

each parameter is indicated the values considered

and if is a default value or a user-specified value

(input).

Table 7: Techniques configurations for approach B.

Techni

ue Settin

Value T

Minrec Node 10 In

Max De

th 7 In

Min

ct S

lit 0.1 In

urit

Metric Gini In

Minrec S

lit 20 In

Min

ct Node 0.05 In

Pre

Auto On In

Pairwise Threshol

0 In

Sin

leton Threshol

0 In

SVM

Conv tolerance 0.001 In

Active Learnin

Enable In

Kernel Function Linea

Default

Com

lexit

facto

0.142831 Default

Pre

auto On In

3.4 Results

3.4.1 Approach A

The results obtained by approach A were very poor.

The best model presents 73.27% of overall accuracy.

To evaluate these models only was used the

accuracy metric, the sensitivity and specificity

metrics were not calculated because the outputs are

represented by more than two classes. Table 8

presents the three best models for this approach.

Table 8: Three best models for approach A.

Scenario Target Technique Accuracy

S1 T2 NB 73.28%

S2 T1 NB 73.28%

S2 T4 NB 50.75%

3.4.2 Approach B

To evaluate the results obtained by approach B were

considered three metrics: sensitivity, accuracy and

specificity. As mentioned before sensitivity is the

most adequate metric in the medicine area.

Table 9 presents the three best models. Table 10

presents the best models for each metric. This table

is useful to find which is the better model when it is

Table 9: Three best models for approach B.

Scenario Technique Accuracy Sensitivity Specificity

S7 DT 74.620%

87.322%

71.169%

S13 DT 77.286% 78.355% 76.628%

S11 DT 77.063% 78.673% 76.628%

expected an equilibrated prediction (accuracy),

predict hourly discharge (sensitivity) or predict a

non-discharge hour (specificity).

Table 10: Three best models for each metric - approach B.

Accuracy Sensitivity Specificity

S2DT - 77.311% S4SVM - 96.140% S12DT - 77.067%

S7DT - 77.286% S11SVM - 95.963% S8DT - 77.043%

S11DT - 77.063% S7SVM - 95.902% S9DT - 77.043%

To understand the importance of each attribute, a

ranking was made for the best model (Table 11).

Minimum Description Length, a supervised

technique for calculate the attribute importance, was

used.

Table 11: Attribute importance for the best model of

approach B.

Attribute Rank Importance

Hour 1 0.110

EC_AC_TOT 2 0.043

EC_AC_HR 3 0.025

ADMINF 4 0.020

EC_AC_O2 5 0.019

Age 6 0.010

Cardio 7 0.008

Renal 8 0.002

ADMINT 9 0.001

Considering scenario 7 and the DT technique,

the most important attributes correspond to the new

variables introduced: number of inpatient hours

(Hour), the hourly accumulated critical events

(EC_AC) for heart rate and oxygen saturation, the

total of accumulated critical events by hour and the

place where the patient was previously hospitalized

(ADMINF).

4 DISCUSSION

The models which presented the best results were

the ones that predict the patient discharge for the

next hour and were able to update the LOS in real-

time. The models of approach A, which predicted

the LOS using information of the admission process

and some measures of patient revealed to be very

poor. In the context of medicine these models cannot

be considered. DT presented overall accuracies

ranging from 7.6% to 38.46%. For NB the results

were situated between 15.14% and 73.28%. In the

case of SVM the worst result corresponds to an

accuracy of 32.65% and the better one to 47.11%.

The use of different classes as target does not

KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing

252

influence the results (the discretization techniques

have no effect in the behaviour of the models). As

mentioned before, one of the most important

measures for evaluating data mining models in the

clinical area is the sensitivity. In the approach A this

measure cannot be evaluated because the target

variable has more than two classes.

Finalizing, the approach A does not evidence

clinical relevance since the corresponding accuracies

are very low and does not achieved the baseline

defined in accordance with physicians (accuracy >

75%). For approach B the models which present best

results are the ones that consider case mix, SOFA,

ACE and the second group of ratios (R2). The model

that uses Scenario 7 and DT technique presented a

good overall accuracy of 74.620% and a very good

sensitivity 87.322%. This model is equilibrated

when compared all the metrics. It revels to be a good

model to make an accurate prediction of patient

discharge. However and being sensitivity one of the

most important metrics there are some models which

presented sensitivities upper than 95%. In this case

the scenario 4 using SVM it is the best model,

presenting a sensitivity of 96.140%. In Figure 4 it is

possible to observe the ROC for the best model.

Figure 4: ROC Curve for model S7-DT.

The use of discrete values for ACE and ratios

produced models with better accuracies.

Nevertheless, sensitivity decreased about 15% to

20%. Although the efforts made during this work,

the accuracy level attained is not very high. The

models with better accuracy presented a lower

sensitivity. So the choose of the best model depends

on the decision-maker objective: good models to

predict the possibility of a patient be discharged

(sensitivity) or equilibrated models between the two

outputs (accuracy). The SVM technique presented

very good results in terms of sensitivity but the

overall accuracy of these models is very weak. The

specificity maintains almost the same results

independently the scenario or technique used.

5 CONCLUSIONS AND FUTURE

WORK

The approaches presented in this work differ in the

data set used and in the number classes of the target

variable. Approach A considered the worst values

occurred during the first 24 hours to predict a

number of days (LOS) that ranges from 1 to n (n

classes). Approach B considered real-time values

hourly collected to predict if a patient will be

discharged in the next hour (2 classes). Approach A

corresponded to a more conventional approach

followed in other areas of medicine. In intensive

medicine such approach is not useful. The first part

of this work was dedicated to approach A. In the

second part, an alternative approach has been

explored in order to overcome the limitations

encountered in approach A. Approach B revealed

characteristics more appropriate to intensive

medicine. All the work was carried out under the

supervision off ICU physicians. They verified the

soundness of the work and the impact of the results

in intensive medicine. By the way of this research

work interesting results were obtained for predicting

LOS in Intensive Medicine. The results allow

conclude that the prediction of patient discharge in

an hourly base is very efficient. Models induced to

predict LOS using admission information and

excluding ratios and attributes from the SOFA reveal

to be very limited. The discretization of some

variables help to increase the overall accuracy of the

models however the sensitivity decreases. This

approach could not be considered the best due the

results presented are not satisfactory.

Intensive Medicine has specific characteristics

that make classic LOS models inadequate. Real-time

models, able to update the LOS according to the

patient condition, obtained better results.

This was an applied research work whose main

contributions are:

 Intensive medicine – an efficient predictive

model to determine LOS. According to the

literature, completely new results were

presented;

 Data mining – a real-time approach has been

developed. Results are an interesting

increment in the scientific knowledge of

streaming data mining.

Those contributions are original and mainly

focused in the field of Intensive Medicine. By using

Data Mining techniques has been possible achieve a

LOS predictive model to improve the discharge

decision process. This approach can be followed in

Real-TimeDataMiningModelsforPredictingLengthofStayinIntensiveCareUnits

253

other critical units where the patients are in

continuous monitoring. In the other cases, solutions

presented in the literature can achieve this goal.

In the future additional variables will be

considered to understand how they can affect the

LOS of inpatients.

ACKNOWLEDGEMENTS

This work has been supported by FCT – Fundação

para a Ciência e Tecnologia in the scope of the

project: Pest-OE/EEI/UI0319/2014.

The authors would like to thank FCT

(Foundation of Science and Technology, Portugal)

for the financial support through the contract

PTDC/EEI-SII/1302/2012 (INTCare II).

REFERENCES

Azari, A., Janeja, V.P. & Mohseni, A., 2012. Predicting

Hospital Length of Stay (PHLOS): A Multi-tiered

Data Mining Approach. In 2012 IEEE 12th ICDB

Workshops. IEEE, pp. 17–24.

Breiman, L. et al., 1984. Classification and Regression

Trees no, ed., Belmont, CA: Wadsworth International

Group.

Caetano, N. et al., 2014. A Data-Driven Approach to

Predict Hospital Length of Stay: A Portuguese Case

Study. In ICEIS 2014. (Accepted to be published)

Cios, K. et al., 2007. Supervised Learning: Statistical

Methods. In A Knowledge Discovery Approach.

Springer US, pp. 307–386.

Clarke, A., 1996. Why are we trying to reduce length of

stay? Evaluation of the costs and benefits of reducing

time in hospital must start from the objectives that

govern change. Quality in health care: QHC, 5(3),

pp.172–9.

Dahan, H. et al., 2014. Proactive Data Mining with

Decision Trees, New York, NY: Springer New York.

Guy W., Rush J, e. a. (2000). Clinical Global Impressions

(CGI) Scale. Psychiatric Measures, APA.

Hachesu, P.R. et al., 2013. Use of data mining techniques

to determine and predict length of stay of cardiac

patients. Healthcare informatics research, 19(2),

pp.121–129.

Han, J., Kamber, M. & Pei, J., 2011. Data Mining:

Concepts and Techniques 3rd ed., San Francisco, CA,

USA: Morgan Kaufmann Publishers Inc.

Hand, D.J., Smyth, P. & Mannila, H., 2001. Principles of

data mining. Available at: http://dl.acm.org/citation.

cfm?id=500820 [Accessed May 5, 2014].

Isken, M.W. & Rajagopalan, B., 2002. Data Mining to

Support Simulation Modeling of Patient Flow in

Hospitals. Journal of Medical Systems, 26(2), pp.179–

197.

Kantardzic, M., 2011. Data Mining: Concepts, Models,

Methods, and Algorithms 2nd ed., Wiley-IEEE Press.

Marshall, A., Vasilakis, C. & El-Darzi, E., 2005. Length

of Stay-Based Patient Flow Models: Recent

Developments and Future Directions. Health Care

Management Science, 8(3), pp.213–220.

Michie, D., Spiegelhalter, D.J., et al., 1994. Machine

Learning, Neural and Statistical Classification. Ellis

Horwood.

Portela, F., Santos, M.F., Silva, Á., et al., 2013. Data

Mining for Real-Time Intelligent Decision Support

System in Intensive Care Medicine. In ICAART 2013

- International Conference on Agents and Artificial

Intelligence. pp. 270–276.

Portela, F., Santos, M.F., Machado, J., et al., 2013.

Pervasive and Intelligent Decision Support in Critical

Health Care Using Ensembles. In ITBAM. pp. 1–16.

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

Predictive Models for Pervasive Intelligent Decision

Support in Intensive Care Medicine. In KMIS. pp. 81–

88.

Quiland, J.R., 1986. Induction of Decision Tress. Machine

Learning, 1(1), pp.81.106.

Rokach, L. & Maimon, O., 2010. Supervised Learning. In

O. Maimon & L. Rokach, eds. Data Mining and

Knowledge Discovery Handbook. Springer US, pp.

133–147.

Santos, M.F. & Azevedo, C., 2005. Data Mining

Descoberta de conhecimento em base de dados, FCA -

Editora de Informática, Lda.

Santos, M.F., Portela, F. & Vilas-Boas, M., 2011.

INTCARE : multi-agent approach for real-time

intelligent decision support in intensive medicine.

Taylor, K., 2010. Oracle Data Mining Concepts.

Vapnik, V.N., 1995. The Nature of Statistical Learning

Theory 1st ed., Springer.

Wagener, G. et al., 2011. The Surgical Procedure

Assessment (SPA) score predicts intensive care unit

length of stay after cardiac surgery. The Journal of

thoracic and cardiovascular surgery, 142(2), pp.443–

50.

Wirth, R., & Hipp, J. (2000). CRISP-DM: Towards a

standard process model for data mining. Paper

presented at the 4th ICPAKDD.

Zhang, X.-C., Zhang, Z.-D. & Huang, D.-S., 2012.

Prediction of length of ICU stay using data-mining

techniques: an example of old critically Ill

postoperative gastric cancer patients. Asian Pacific

journal of cancer prevention: APJCP, 13(1), pp.97–

101.

KMIS2014-InternationalConferenceonKnowledgeManagementandInformationSharing

254