HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME

IN INTENSIVE CARE BASED ON DATA MINING TECHNIQUES

Marta Vilas-Boas, Manuel Filipe Santos, Filipe Portela

Departamento de Sistemas de Informação, Universidade do Minho, Guimarães, Portugal

Álvaro Silva, Fernando Rua

Serviço de Cuidados Intensivos, Hospital Geral de Santo António, Porto, Portugal

Keywords: INTCare, Intelligent Decision Support Systems, Clinical Data Mining, Real-time prediction, Hourly

prediction, Intensive Care Medicine.

Abstract: The use of Data Mining techniques makes possible to extract knowledge from high volumes of data.

Currently, there is a trend to use Data Mining models in the perspective of intensive care to support

physicians’ decision process. Previous results used offline data for the predicting organ failure and outcome

for the next day. This paper presents the INTCare system and the recently generated Data Mining models.

Advances in INTCare led to a new goal, prediction of organ failure and outcome for the next hour with data

collected in real-time in the Intensive Care Unit of Hospital Geral de Santo António, Porto, Portugal. This

experiment used Artificial Neural Networks, Decisions Trees, Logistic Regression and Ensemble Methods

and we have achieved very interesting results, having proven that it is possible to use real-time data from the

Intensive Care Unit to make highly accurate predictions for the next hour. This is a great advance in terms

of intensive care, since predicting organ failure and outcome on an hourly basis will allow intensivists to

have a faster and pro-active attitude in order to avoid or reverse organ failure.

1 INTRODUCTION

In the information era, Intensive Care Units (ICU)

are a particularly attractive area for the use of Data

Mining (DM) techniques. Large quantities of clinical

data is produced and stored every day. However, the

databases may have valuable unknown information

regarding the patient’s prognosis. Human medical

data is the most rewarding and complicated of all

biological data to mine and analyze (Cios, 2002).

Critical patients’ condition is so complex that

sometimes even doctors find it hard to decide about

the most adequate procedure to provide them the

best health care possible. Despite of the patient’s

vital signs being constantly monitored, they only

serve as alarms to inform when a patient’s condition

is deteriorated; they don’t have the capability of

predicting those conditions, leading to a reactive

attitude of the medical staff. Subsequently, the

challenge is to use DM techniques to discover

unknown patterns and to predict dysfunction and

organ failure, as well as the patient’s outcome in a

timely manner, so that physicians may have a pro-

active attitude towards the patients’ best interest.

Hence the new approach of hourly prediction.

The use of DM techniques in the medical arena has

been gaining an increasing interest by researchers

but, despite the high expectatives, its application in

real world settings has been limited (Bellazzi, 2008).

To fill this gap, we are developing the INTCare

system (Santos, 2006), (Gago, 2006), (Silva, 2006),

an Intelligent Decision Support System (IDSS)

(Gago, 2008) that makes uses of DM techniques for

predicting organ failure and outcome for the next

hour. INTCare (PTDC/EIA/72819/2006) has

evolved greatly and it is currently being tested in

Hospital Geral de santo António (HGSA), Porto,

portugal.

The purpose of this paper is to present the new DM

models generated with data collected in the ICU of

HGSA, in real-time for hourly prediction of organ

failure and outcome.

270

Vilas-Boas M., Filipe Santos M., Portela F., Silva Á. and Rua F. (2010).

HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME IN INTENSIVE CARE BASED ON DATA MINING TECHNIQUES.

In Proceedings of the 12th International Conference on Enterprise Information Systems - Artiﬁcial Intelligence and Decision Support Systems, pages

270-277

DOI: 10.5220/0002903802700277

 SciTePress

Section 2 presents an overview on prediction of

organ failure and outcome and previous work, as

well as the INTCare system. Section 3 describes the

data modelling for the creation of the DM models

for each organ system and outcome. Later, it

presents the results, whose predictive performance is

discussed in section 4. Section 5 concludes this

paper with some considerations regarding this study

and it is pointed the future work.

2 BACKGROUND AND RELATED

WORK

Throughout the past 30 years, clinical scores have

been used to assess severity of illness and organ

failure in ICU and to stratify patients according to

their clinical condition (Silva, 2008). Prognostic

scores have undergone significant development,

validation, and refinement. Several models of organ

failure and mortality risk prediction have become

available, including the Sequential Organ Failure

Assessment (SOFA), which scores six organ systems

(cardiovascular, respiratory, renal, coagulation, liver

and neurological) (Le Gall, 2005). However, the

probabilistic nature of those models compromises

their application to the individual patient and is not

universally accepted (Silva, 2007). Each patient has

its individual characteristics; therefore, apart from a

global learning over patients, a more local learning

of would be useful (Ramon, 2007).

Recent innovations in risk adjustment include

automatic physiologic and diagnostic variable

retrieval and the use of artificial intelligence

(Rosenberg, 2002). Assessing the degree of organ

failure is a crucial task in ICU since one of the

critical aspects of ICU is to avoid or reverse organ

failure (Vincent, 1996). The earlier the patient’s risk

is identified, the earlier a correct therapy can be

applied.

DM is an important step of a process entitled

Knowledge Discovery in Databases (KDD) (Fayyad,

1996) which has brought the researchers’ attention.

The scope of DM goes beyond complex queries

to databases for hypothesis validation; it also

includes the discovery of new and previous

unknown patterns (Fayyad, 1998).

DM techniques can be applied in the

development of prognostic models to be integrated

in a Decision Support System (DSS). Some attempts

have been made to predict organ failure (Toma,

2008) but, to our knowledge, there is no DSS

implemented in a real environment that uses online

and real time data to predict organ failure for the

next hour.

Next, it is presented the INTCare system, an

Intelligent Decision Support System (IDSS) that

uses DM techniques for prediction of organ failure

and outcome, in real-time, for the next hour.

2.1 The INTCare System

INTCare is a situated IDSS for intensive medicine

that is being developed in the ICU of the HGSA in

Porto, Portugal. It relies on intelligent agents

(Abelha, 2004) that perform autonomous actions in

order to meet its purposes (Gago, 2006), (Santos,

2006). It is capable of predicting organ failure of six

systems (cardiovascular, respiratory, renal, liver,

coagulation and neurological) and outcome, i.e. the

patient’s status at the time of discharge (dead or

alive). Although the results attained could be

considered as promising (Gago, 2008), the previous

models were developed using offline learning, with

data from 42 European ICU (Miranda, 1999). The

challenges are to further develop these models so

that they can be used throughout the ICU. They are

part of the DM agent of the INTCare system, which

is responsible for the retrieval of relevant data in

order to make possible the application of the DM

models in real time (Gago, 2006). This system is

semi-autonomous, avoiding the need for manual data

preparation and integration, which has been pointed

as a great obstacle in the implementation of

prediction models as a decision support tool

(Ramon, 2007), (Patel, 2008).

As pointed before by research on the INTCare

(Gago, 2008), the new developments should regard

the inclusion of the data available as it is registered

(i.e. online data) to guarantee that all relevant

clinical information is taken into account for

decision support in real-time.

This is possible due to the new improvements of

the INTCare system that uses, not only the data

gathered by the bedside monitors (e.g. heart rate),

but also data from the Electronic Nursing Record

(ENR) (e.g. lab results) (Santos, 2009).

3 PREDICTION MODELS

FOR THE NEXT HOUR

3.1 Data Description

The data used to generate the DM models was

gathered in the ICU of HGSA during the period

HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME IN INTENSIVE CARE BASED ON DATA MINING

TECHNIQUES

271

between January 1

and March 31

of 2009 and it

relates to the first five days of stay of thirty two

patients.

The data collected came from three distinct

sources: the Electronic Health Record (EHR), 10

bedside monitors and paper-based nursing records.

The input variables consist of the case mix (i.e.

age, admission type and admission from) and

Critical Events (CE), derived from four physiologic

variables collected by the bedside monitors (Blood

pressure, heart rate and oxygen saturation) and urine

output, registered on the nursing records. CE was

defined by a panel of experts (Silva, 2003).

Whenever a physiological parameter is out of its

normal range for more than 10 minutes, it is

considered a CE.

The output target considered is the organ status

(normal and dysfunction/failure) of six organ

systems (cardiovascular, respiratory, coagulation,

hepatic, neurological and renal), measured by the

SOFA score and the patient’s outcome (dead or

alive).

The new requirements of the INTCare system

encompass a finer grained prediction, i.e. the

prediction for the next hour, hence the need for

hourly data. In order to accomplish that, some

adjustments were made.

The number of CE of each physiological

parameter considered was calculated hourly, for

each patient, and subsequently Accumulated Critical

Events (ACE) was derived as a new variable. The

CE were calculated with data from the bedside

monitors (which registers the patients’ vital signs at

the frequency of one minute) and from the paper-

based nursing record, in the case of CE of urine

output. The SOFA scores aren’t usually registered,

so they were calculated manually with data from the

bedside monitors and the paper-based nursing

record. Generally, SOFA is calculated once a day

and we considered the same score for each hour of

the corresponding day, except when there was data

to calculate it more than once a day. Later, it was

adapted to a binary representation, where 0 means

normality and 1 means dysfunction/failure and

comprises SOFA scores of 1,2,3,4.

Figure 1 shows the distribution of the values of

the target variables, as well as its missing values.

Noisy data (incorrect values) was manually

detected and corrected by ignoring values

considered absurd by the medical experts. This led

to an increase of missing data.

All the data considered was integrated in a

database for the construction of the final dataset for

the models’ creation, with 2614 records and 20

variables.

Figure 1: Distribution of the values for the target

variables (%).

3.2 Modelling (Feature Selection)

For the models’ creation, we had two concerns, the

first being related to the features selection and the

second, the DM techniques used.

We have explored three scenarios (M1, M2 and

M3) regarding the variables to be included in the

models, as shown below:

M1 = {Hour, Case Mix, ACE}

M2 = {Hour, Case Mix, ACE, Ratios}

M3 = {Hour, Case Mix, ACE, SOFA}

Where

Case Mix = {Age, Admission type, Admission

from}

ACE = {ACE of Blood Pressure, ACE of Oxygen

Saturation, ACE of Heart Rate, ACE of Urine

Output}

Ratios = {ACE of BP/elapsed time of stay, ACE of

SO2/elapsed time of stay, ACE of HR/elapsed time

of stay, ACE of Ur/elapsed time of stay}.

All these variables have hourly values.

By exploring these scenarios we intend to attest

the importance of new variables, other than the CE

for the prediction. The inclusion of ratios points to

the severity of the patient’s clinical evolution and its

relation to organ failure and outcome. Since multiple

organ failure is a major cause of ICU mortality

(Amaral, 2005), it is justified the inclusion of the

SOFA scores of the other organ systems.

Due to biased distribution of the target variables,

as shown in Figure 1, we transformed the variables

using the logarithmic function in order to maximize

normality and avoid overfitting.

For each one of these scenarios and for each

target variable (cardiovascular, respiratory, renal,

coagulation, liver and outcome) were applied four

DM techniques: Artificial Neural Networks (ANN),

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

272

Decision Trees (DT), Regression and Ensemble

methods. These techniques have been applied before

for the next day prediction and with different data.

For the ANN, several configurations were tested

and the best results were achieved with a fully

connected Multilayer Perceptron with 3 hidden

neurons and logistic activation function. To assure

statistical significance, 30 runs were applied to all

tests.

Similarly to the ANN, for the DT, various

configurations were tested in order to achieve the

best results. The splitting method used for

partitioning the data was the Gini reduction. The

default algorithm splits a node into 2 branches and,

to avoid overfitting, the maximum number of

branches from a node was set to 10 and the splits

were evaluated as a reduction in impurity (Gini

index). The maximum depth of the tree was 6.

For the Regression, since we have binary targets,

it was applied a logistic regression. The selection

method that produced best results was the Stepwise.

The ensemble method used a combined mode of

the ANN, DT and Regression with the Mean

probability function.

It was not possible to generate models for the

neurological system due to the amount of missing

data (83%) required for calculating its SOFA score.

3.3 Results

A total of 6 (target variables) * 3 (scenarios) * 4

(DM techniques) = 72 models were generated and

tested.

We have partitioned the dataset in learning,

validation and test subsets to objectively assess the

predictive performance. The partition method is

based on a holdout of 40% for training, 30% for

validation and 30% for test.

Next, we present the results obtained for each

organ system (Tables 1 to 5) and outcome (Table 6),

in terms of Accuracy (Acc), Specificity (Spec) and

Sensibility (Sens) of each technique and scenario.

Table 1: Results of the Cardiovascular System (%).

M Technique Acc Spec Sens

Reg. 78.2 69.5 79.6

ANN 84.2 72.4 88.3

DT 87.6 77.5 92.4

Ens. 71.4 44.6 74.5

Reg. 81.4 79.8 81.7

ANN 91.9 93.7 91.3

DT 88.4 81.8 90.6

Ens. 63.3 30.2 73.3

Reg. 85.1 86.5 85

ANN 92.7 90.6 93.4

DT 88.6 81.6 90.9

Ens. 61.8 31.1 73.9

Table 2: Results of the Respiratory System (%).

M Technique Acc Spec Sens

Reg. 92 93.3 84.2

ANN 95.1 96.3 89.2

DT 95.6 79.6 87.3

Ens. 85.3 87.8 65.1

Reg. 92.4 94 82.7

ANN 96.1 96 96.2

DT 91.3 94.4 75

Ens. 72.9 84.9 22.8

Reg. 96.8 97.5 93.9

ANN 98.1 99 94.4

DT 95.8 96.1 95

Ens. 75.2 85.5 32.8

Table 3: Results of the Renal System (%).

M Technique Acc Spec Sens

Reg. 86.3 84.5 86.7

ANN 96.7 93.8 97.7

DT 93.2 93.4 93.1

Ens. 72.5 46.5 81.2

Reg. 86.1 75.2 89

ANN 95.4 87.7 98

DT 93.9 92.5 94.3

Ens. 67.9 35.2 79.7

Reg. 93.7 90.7 94.7

ANN 96.9 93.9 98.1

DT 95.7 92.4 96.8

Ens. 80 62.4 85.6

Table 4: Results of the Coagulation System (%).

M Technique Acc Spec Sens

Reg. 80.9 83.6 76.8

ANN 90.8 88.9 94.2

DT 89.1 96.2 81

Ens. 62.3 68 53.1

Reg. 81.6 84.6 71.2

ANN 98.1 98.5 97.5

DT 91 91.8 89.5

Ens. 56 62.6 45.6

Reg. 87.5 87.1 87.9

ANN 95.3 95.1 95.6

DT 92.1 96.4 86.6

Ens. 64.8 70.1 56.3

Table 5: Results of the Liver System (%).

M Technique Accuracy Specificity Sensibility

Reg. 80 71.2 82.3

ANN 91.7 87.7 93.1

DT 91.2 83.7 94.3

Ens. 59.5 32.1 72.7

Reg. 83.3 72.4 86.8

ANN 93.6 88.5 95.4

DT 93.2 88.3 95

Ens. 64.1 34.8 75.6

Reg. 94.3 95.4 94.1

ANN 98.2 97.6 98.3

DT 96.3 96.9 96.1

Ens. 77.8 61 85.1

HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME IN INTENSIVE CARE BASED ON DATA MINING

TECHNIQUES

273

Table 6: Results of the Outcome (%).

M Technique Acc Spec Sens

Reg. 88.8 83.9 90

ANN 97.3 94.6 98.3

DT 94.5 92.9 95

Ens. 65.4 34.7 81.6

Reg. 89.3 84.3 90.3

ANN 97.4 96 98

DT 94.5 92.5 95

Ens. 63.6 28.2 79.3

Reg. 91.2 87.4 92.1

ANN 97.1 95.1 97.6

DT 94.8 90.6 96.1

Ens. 63.2 31 79.6

4 DISCUSSION

For the analysis of the models’ performance, we

used confusion matrixes, a tool commonly used in

domains where the cost of incorrect classification is

high. In medicine, it is usual to use sensitivity and

specificity analysis for measuring the rates of error

(Cios, 2002). From the matrixes, three measures

were derived. Table 7 synthesises the best results for

accuracy, specificity and sensibility for each organ

system and outcome regarding the four techniques

and the three scenarios considered in the

experiments and presented in Tables 1 to 6.

Table 7: Best results of the organ systems and outcome (%).

Acc Spec Sens

Cardiovascular 92.7

ANN M3

93.7

ANN M2

93.4

ANN M3

Respiratory 98.1

ANN M3

96.2

ANN M2

Renal 96.9

ANN M3

93.3

ANN M3

98.1

ANN M3

Coagulation 98.1

ANN M2

98.5

ANN M3

97.5

ANN M2

Liver 98.2

ANN M3

97.6

ANN M3

98.3

ANN M3

Outcome 97.4

ANN M2

95.1

ANN M3

98.3

ANN M1

As shown in Table 7, and in a global

appreciation, it is notorious the distinction of the

ANN as the technique with best results for the three

metrics and M3 as the scenario with the most

number of best results, followed by M2. However,

as shown in Tables 2 to 6 and Figures 2 to 7, the DT

has a high performance as well. The ensembles show

a very poor performance, whose curves relate to a

random classifier, as shown in Figures 2 to 7.

In the context of clinical decision, for the organ

systems prediction and considering that 0 is the

normality and 1 is the dysfunction/failure, the

assessment must be in favour of sensibility, as it

measures the proportion of positives that are

correctly identified. The same principle was applied

for the prediction of the outcome. Thus, Table 8

presents the best results for the prediction of the

organ systems and outcome and the corresponding

techniques and scenario, in terms of sensibility.

Table 8: Sensibility for the organ systems and outcome (%).

System Technique Scenario Sensibility

Cardiovascular ANN M3 93.4

Respiratory ANN M2 96.2

Renal ANN M3 98.1

Coagulation ANN M2 97.5

Liver ANN M3 98.3

Outcome ANN M1 98.3

In the appendix are represented the ROC curves

for target 1 (dysfunction/failure and death) of the

bests models in terms of sensibility for the 5 organ

systems and outcome. ROC curves are frequently

used in the medical area to evaluate computational

models for decision support, diagnosis and prognosis

(Lasco, 2005). They measure the degree of

discrimination that can be obtained from a given

model and they establish the relation between

specificity and sensibility of a classifier.

As shown in Figures 2 to 7, it is confirmed the

emphasis of the ANN. Nevertheless, for the

respiratory, renal and liver systems (Figures 3, 4, 6),

the DT and regression have a close performance to

the ANN.

From Figures 2 to 7 it is also notorious the poor

performance of the ensemble methods, which point

to random classifiers, except for the renal and liver

systems, whose performance is slightly better, but

nonetheless poor.

With these experiments, it is confirmed the

importance of the inclusion of CE in organ failure

prediction. However, the inclusion of new variables

such as the ratios and SOFA scores of other organ

systems generated models with better results. Half of

the best models (Cardiovascular, Renal and Liver)

are related to the third scenario, hence it is justified

the new approach of including the SOFA score of

the other systems, as well as the ratios.

For the outcome prediction, the best models were

achieved with the first scenario, which only includes

the CE as input variables.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

274

5 CONCLUSIONS

AND FURTHER WORK

Currently, in the ICU environment, the decision

making is based on severity scores like SOFA and in

the intensivists’ empirical experience. However, this

process has some limitations. It is based on

probabilistic scores and doesn’t take into account

import information that might be hidden regarding

the patient’s clinical status and its evolution (e.g.

bedside monitored physiological parameters).

Research has evolved towards the inclusion of

new variables and DM techniques in order to predict

organ dysfunction/failure and outcome.

This paper presented new DM models for the

INTCare system, with data gathered in real-time at

the ICU of HGSA, in Porto, Portugal. The

techniques tested have been used before in the

research of the INTCare. The novelty of this

approach relies on the use of real-time data and a

finer grained prediction, i.e. the prediction of organ

failure and outcome for the next hour. Moreover,

new variables were used for the models generation

(e.g. racios). The experiments carried out focused on

finding the best combination of techniques and

features selection. For each technique, several

configurations were tested in order to achieve the

best results possible.

In this paper, we have proven that it is possible

to use online and real-time data to generate highly

accurate models, which is a great advance in the

context of a situated intelligent decision support

system like INTCare. The main goal in ICU is to

avoid or reverse organ failure. We expect that, with

an hourly prediction of the patients’ clinical status, it

will be possible for intensivists to have a timely

intervention so that worst complications may be

avoided.

As pointed in Section 3, the data used came from

three sources, including a paper-based nursing

record. However, we are developing an ENR that

collects hourly and in real-time all the necessary

data to feed the DM models (Santos, 2009).

These models must be tested in the ICU with

new data in real-time and new models will be

generated in order to try to achieve the best results

possible.

Advances in the prediction of organ failure and

outcome might include the development of new

strategies presented in this paper regarding the

features selection (e.g., the inclusion of new

variables like the SOFA score of the other systems,

and ratios). Moreover, it should be pointed that the

effort was put into the prediction of each organ

system individually and not the systems altogether.

Hence, it is expected that different systems have

different variables and techniques for their

prediction.

Although we have achieved very good results, it

should be stressed out that the dataset used was a

small population of patients. We believe that it will

be possible to maintain a good degree of prediction

because it is known that clinical data is very stable

(Silva, 2007).

The next step is the deployment of the models

and the analysis of their impact on the ICU

environment.

REFERENCES

Abelha, A., Machado, J., Alves, V., Neves, J., 2004. Data

warehousing through multi-agent systems in the

medical arena. In 1st international Conference on

Knowledge Engineering and Decision Support. Porto,

Portugal.

Amaral, A.C., Andrade, F.M., Moreno, R., 2005. Use of

the Sequential Organ Failure Assessment score as a

severity score. In Intensive Care Med, 31, pp. 243-249.

Bellazzi, R., Zupan, B., 2008. Predictive Data Mining in

clinical medicine: current issues. In Int J Med Inform,

pp. 81-97.

Cios, K., Moore, G., 2002. Uniqueness of medical data

mining. In Artificial Intelligence in Medicine 26, 1–24.

Fayyad, U., 1998. Mining databases: Towards algorithms

for knowledge discovery. In IEEE Data Eng. Bull., pp.

39-48.

Fayyad, U., Piatetsky-Shapiro, G., Smyth, P., 1996. The

KDD process for extracting useful knowledge from

volumes of data. In Communications of the ACM, 39,

pp. 27-34.

Gago, P., Santos, M. F., 2008. Towards an intelligent

decision support system for intensive care units. In

The 18th european conference on artificial

intelligence.

Gago, P., Santos, M. F., Silva, Á., Neves, J., Gomes, L.,

2006. INTCare: A knowledge discovery based

intelligent decision support system for intensive care

medicine: an overview. In Journal of decision

systems, pp. 241-259.

Lasco, T., Bhagwat, J., Zou, K., Ohno-Machado, L., 2005.

The use of receiver operating characteristic curves in

biomedical informatics. In J. of Biomedical

Informatics, 38(5):404–415.

Le Gall, J., 2005. The use of severity scores in the

intensive care unit. In Intens Care Med 2005;31:1618-

23.

Miranda, R., 1998. The effect of harmonizing and

standardizing the nursing tasks on intensive care units

of the European Community, second European

intensive care unit study (EURICUS-II). In Technical

report BMH4-CT96-0817, Health services research

unit – Groningen university hospital, The Netherlands.

HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME IN INTENSIVE CARE BASED ON DATA MINING

TECHNIQUES

275

Patel, V., Shortliffe, E., Szolovits, P., 2008. The coming of

age of artificial intelligence in medicine. In Artif Intell

Med.

Ramon, J., Fierens, D., Guiza, F., Meyfroidt, G., Blockeel,

H., Bruynooghe, M., Berghe, G., 2007. Mining data

from intensive care patients. In Advanced Engineering

Informatics, Elsevier Science Publishers, pp. 243-256.

Rosenberg, A.L, 2002. Recent innovations in the intensive

care unit risk-prediction models. In Current Opinion

in Critical Care, 8, pp. 321-330.

Santos, M.F., 2006. Intelligent decision support in

Intensive Care Medicine. In 2nd International

Conference on Knowledge Engineering and Decision

Support, Lisbon, Portugal, pp. 401-405.

Santos, M.F., Portela, F., Vilas-Boas, M., Machado, J.,

Abelha, A., Neves, J., 2009. Intelligent Decision

Support in Intensive Care Units – Nursing Information

Requirements. In Book Chapter on Applied

Informatics and Communications (AIC’09).

Silva, Á., 2007. Modelos de inteligência artificial na

análise da monitorização de eventos clínicos adversos,

disfunção/falência de órgãos e prognóstico do doente

crítico. Porto. In Tese de doutoramento.

Silva, Á., Cortez, A., Santos, M.F., Gomes, L., Neves, J.,

2008. Rating organ failure via adverse events using

Data Mining in the intensive care unit. In Artificial

intelligence in medicine, pp. 179-193.

Silva, Á., Cortez, P., Santos, M.F., Gomes, L., Neves, J.,

2006. Mortality assessment in intensive care units via

adverse events using artificial neural networks. In

Artificial Intelligence in Medicine, pp. 223-234.

Silva, Á., Santos, P., Gomes, L., Neves, J., 2003. Organ

failure diagnosis by artificial neural networks. In Third

IASTED International Conference on Artificial

Intelligence and Applications. Málaga.

Toma, T., Abu-Hanna, A., Bosman, R., 2008. Discovery

and integration of univariate patterns from daily

individual organ-failure scores for intensive care

mortality prediction. In Artificial Intelligence in

Medicine, pp. 47-70.

Vincent, J., Moreno, R., Takala, J., Willastss, S.,

Mendonca, A., Bruining, H., 1996. The SOFA

(Sepsis-related Organ Failure Assessment) score to

describe organ organ dysfunction/failure. In Intens

Care Med 1996;22:707-10.

APPENDIX

Figure 2: ROC curve of the Cardiovascular system for

target 1 (dysfunction/failure) of M3.

Figure 3: ROC curve of the Respiratory system for target 1

(dysfunction/failure) of M2.

Figure 4: ROC curve of the Renal system for target 1

(dysfunction/failure) of M3.

ICEIS 2010 - 12th International Conference on Enterprise Information Systems

276

Figure 5: ROC curve of the Coagulation system for target

1 (dysfunction/failure) of M2.

Figure 6: ROC curve of the Liver system for target 1

(dysfunction/failure) of M3.

Figure 7 - ROC curve of the Outcome for target 1 (dead)

of M1.

HOURLY PREDICTION OF ORGAN FAILURE AND OUTCOME IN INTENSIVE CARE BASED ON DATA MINING

TECHNIQUES

277