Using Domain Knowledge to Improve Intelligent Decision Support in

Intensive Medicine

A Study of Bacteriological Infections

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

Keywords: Antibiotics, Therapies, Infections, Bacteria, Intensive Care Units, Heuristics, Artificial Intelligence,

INTCare, Decision Support.

Abstract: Nowadays antibiotic prescription is object of study in many countries. The rate of prescription varies from

country to country, without being found the reasons that justify those variations. In intensive care units the

number of new infections rising each day is caused by multiple factors like inpatient length of stay, low

defences of the body, chirurgical infections, among others. In order to complement the support of the

decision process about which should be the most efficient antibiotic it was developed a heuristic based in

domain knowledge extracted from biomedical experts. This algorithm is implemented by intelligent agents.

When an alert appear on the presence of a new infection, an agent collects the microbiological results for

cultures, it permits to identify the bacteria, then using the rules it searches for a role of antibiotics that can

be administered to the patient, based on past results. At the end the agent presents to physicians the top-five

sets and the success percentage of each antibiotic. This paper presents the approach proposed and a test with

a particular bacterium using real data provided by an Intensive Care Unit.

1 INTRODUCTION

Infections are the principal cause of mortality in

Intensive Care Units (ICU) both in Europe (Angus et

al., 2001) and USA (Vincent et al., 2006). According

to the medical community this type of problem is

most common in patient with more than five days of

stay.

Vincent et al. (2009) conducted a study about

infections in the ICU. In this study it was evaluated a

set of patient where 50% of the patients had an

infection. From the infected patients, in 70% of the

cases were prescribed therapeutics associated with

antibiotics and microbiological cultures being in

generally obtained positive results.

The antibiotic prescription varies pretty much

from country to country, being the reason unknown

(Lindbaek, 2006).

In many cases the patient body resists to the

administered antibiotic, being necessary to test and

prescribe another antibiotic. The main purpose of

this work is to explore a complementary approach in

order to support the decision process related to

infections. This work aims to answer to the question

“Is it possible to use an intelligent approach in

order to support the infection decision process?”

This paper introduces an algorithm based in the

heuristic concept to provide information about

alternative antibiotics that can have success in

control of a particularly infection.

The algorithm will be embedding in an

intelligent decision support system for intensive

medicine – INTCare system. The system is based

on agents responsible to collect and process the data

in real-time. The heuristic aims for scanning new

infections or patterns of infections. Whenever an

infection is found the system automatically runs the

algorithm returning to the clinical staff a list of

possible antibiotics that can be administered to the

patient.

The algorithm is responsible for finding the best

treatments for a specific infection. The system

searches all the treatments performed in the past and

based on the patient clinical data, admission

582

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

Using Domain Knowledge to Improve Intelligent Decision Support in Intensive Medicine - A Study of Bacteriological Infections .

DOI: 10.5220/0005286405820587

In Proceedings of the International Conference on Agents and Artiﬁcial Intelligence (ICAART-2015), pages 582-587

ISBN: 978-989-758-074-1

 2015 SCITEPRESS (Science and Technology Publications, Lda.)

variables like age and sex, and shows some

alternative treatments that can achieve success in the

combat of the infection. This research also considers

factors like the cost, successful cases and expected

time to the antibiotic produce effect.

The system returns then a set of treatments

successfully applied in the past using input variables

similar to the patient data (case based approach).

To test this system they were used real data

provided by Intensive Care Unit of Centro

Hospitalar do Porto (CHP), Porto, Portugal. All of

this work is inserted in the INTCare research project.

This article is divided into five sections. The first

one consists in the introduction of the paper. The

main concepts and related work are described in

background that is the section number two of this

paper. On the third section it is described the data

analysis, the data transformations and it is presented

the heuristic algorithm and achieved results with the

application of the heuristic. Then in the section four

it is analysed and discussed the results and on the

last section they are presented some considerations

as well the future work.

2 BACKGROUND

2.1 Infection, Bacteria and Antibiotics

in Healthcare

Among bacteria, viruses, fungi and parasites, a very

small amount of these organisms are linked to the

development of infections, usually called pathogens.

An infectious disease is normally characterized by

symptoms like fever, pain and swelling and others

more specific to the organs that the infection is

affecting. The diagnosis of an infection is done

recurring to the isolation of a pathogen and grown of

them in artificial cultures. So these pathogens can be

identified and then it is administrated to the patient

the most convenient antibiotic (Ryan et al., 2004).

In Europe in each year about 4 million of

patients acquire an infection associated to healthcare

and approximately 37000 of them dies from the

contracted infection. The most common infections in

healthcare are: respiratory, urinary, surgical,

gastrointestinal and bloodstream (ECDC, 2012).

Bacteria are microorganisms; the smallest among

the living cells, without nucleus but all of them have

the nucleic acid and protein synthesis (Ray et al.,

2004).

The most common bacteria in healthcare are

Escherichia Coli, Pseudomonas aeruginosa,

Candida and Enterococcus species (ECDC, 2012).

The Pseudomonas aeruginosa is a common bacteria

from the Gram-negative class, it is a common

pathogen in inpatients with a length of stay superior

to one week. The infections provoked by this

pathogen are many and can be located on respiratory

track (pneumonia), heart (endocarditis), bloodstream

(bacteraemia) among others (Ray et al., 2004).

2.2 Heuristics

The term heuristic comes from the Greek and means

“to find out, discover”. Romanycia and Pelletier

(1985) created a survey of definitions and concluded

that a heuristic is a device that is used in problem

solving. It can be a strategy, knowledge, computer

program, data structure and have to guarantee the

supply of a solution. The heuristic developed in this

work is based on the medical domain knowledge

about antibiotics and infections in intensive

medicine.

2.3 INTCare

This study is being developed under the research

project INTCare. INTCare is an Intelligent Decision

Support System (IDSS) that is in constant

development and testing. It is deployed in the ICU

of the Centro Hospitalar do Porto (CHP). This IDSS

is based on intelligent agents (Santos et al., 2011)

and aims to predict clinical events as patient organ

failure (cardiovascular, respiratory, renal, hepatic,

neurological and hematologic systems), possible

readmissions, diseases, infections and patient

outcome (Portela et al., 2013). The system is able to

suggest procedures, treatments and therapies. This

system is based in four subsystems (data acquisition,

knowledge management, inference and interface)

that recur to intelligent agents in order to take

actions (Portela et al., 2013).

3 STUDY DESCRIPTION

3.1 Data Understanding

As above referred the administration of antibiotics is

an important issue because there is a great number of

particularities that should be analysed in the

treatment of infections. The system developed for

this study tries to eliminate some of those issues. In

order to analyse and identify treatment patterns it is

necessary to provide input variables for the system.

After some meetings with the medical staff of the

CHP it was defined a fixed variables package.

UsingDomainKnowledgetoImproveIntelligentDecisionSupportinIntensiveMedicine-AStudyofBacteriological

Infections

583

Independently the type of infection, it should be

always considered four type of variables. They are

age, sex, leucocytes and days of internment. The

remaining variables are related with the organic

system that the infection is affecting: cardiovascular,

respiratory, hepatic, renal, neurological and

hematologic. For the study it was considered

infections provoked by the Pseudomonas aeruginosa

a bacteria that typically affects the respiratory

system. To induce the models they were used four

variables related with this system: SPO2, PCO2 and

PaO2.

Table 1 shows the used variables for this study

on the first part of the heuristic algorithm and a brief

description of each one.

Table 1: Variables Used for Search.

Variable Description

age The patient age in years

sex

Information about the patient sex. Can be

male or female.

leuc

Leucocytes – quantity of leucocytes referred

in the laboratory analysis with the closest

date of the date in where the infection

appears.

doi

Length of stay – the number of days that the

patient is hospitalized in the ICU.

SPO2

Peripheral capillary oxygen saturation – is

an estimation of the oxygen saturation level.

Normal values are considered 95% to 100%.

Between 90% and 95% the oxygen

saturation is low but not necessarily

represents a health issue and below 90% is

considered that a patient is in hypoxemia.

PaCO2

Partial pressure of carbon dioxide –

Represents the partial pressure of CO2 in

arterial blood and expresses the

effectiveness of the alveolar ventilation.

Normal values are between 35 to 45 mm

Hg. If the value is higher than 45 mmHg the

patient is in hypercapnia.

PaO2

Oxygen partial pressure – refers to the

measurement of oxygen in the arterial

blood. The normal range is between 75-100

mm Hg. If the value is below that the

patient is not receiving enough oxygen.

PaO2/FIO2

Ratio of the oxygen partial pressure and

fraction of inspired oxygen – compares the

level of oxygen level in the blood with the

oxygen that is breathed. It’s very useful do

check if exists problems with how lungs

transfer the oxygen to the blood. If the ratio

is below to 250 mm Hg is one of the criteria

of evaluation for pneumonia.

The variables described in table 1 belong to the

search variables group. In order to perform the

heuristic algorithm it is necessary to have a set of

data containing results of past treatments. The

variables represented in this set, are the variables

mentioned in table 1 plus the variables: culture,

antibiotic, antibiotic result, days of treatment and

cost. Table 2 shows the variables present in the set

of data related with the past results.

Table 2: Variables related with past results.

Variable Description

tage, tsex,

tleuc, tdoi,

tSPO2,

tPaCO2,tPaO2,

tPaO2/FIO2

The variables represent the same

results as the ones from Table 1, but

it was added the prefix t for

treatment.

culture

Represents a positive case for some

bacteria. For example if the field as

the text “pseaer” it means the data

from that row represents a positive

case of pseaer bacteria.

antibiotic

This field represents the antibiotic

given to a patient in order to combat

bacteria.

antibiotic

result

The antibiotic result assumes two

results: positive and negative. If the

value is positive it means that the

antibiotic succeeds and the infection

is being stopped. If the value is

negative, it means the antibiotic did

not produce curative effects and it

was necessary to try with another

antibiotic.

Treatment

days

Is the number of days that the

antibiotic was administrated until

makes effect on the combat to the

infection. If the value is null it means

the antibiotic was not suited for the

bacteria.

cost

Represents the cost of the antibiotic

administration. This value is

obtained multiplying the unitary cost

of the dose by the number of days of

treatment.

In order to test the proposed algorithm it was used a

dataset of historic treatments. This dataset only

contains 275 rows referring to 42 inpatients

treatments using 5 different antibiotics. The dataset

size is small because at the moment there are some

laboratory results in a closed format which means

that we do not have access to the results of all the

bacteriological analysis. However to test this

approach the dataset is enough because it represents

a complete sample of an infection.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

584

3.2 Heuristic Approach

The heuristic searches for new infections and

patterns that can indicate an infection. When an

infection is detected it will wait for the results of

cultures analysis. When the culture results denote

the presence of a bacterium a search is performed for

looking for possible treatments.

Next, for a better understanding, an example will

be given using a real data for pseaer bacterium.

Despite of this representation, the heuristic

developed can be used in another type of infection /

bacteria, by only changing the input data and output

target.

An algorithm was developed for searching

possible treatments that best suit the case. The

historic treatments and domain knowledge are used

to find the list of best treatments possible.

The system needs a dataset. These data is

represented by the following groups:

Variables searched = {doi + leuc + age + sex +

SPO2 + PaCO2 + PaO2 + PaO2/FIO2}

Treatment history = {tdoi + tleuc + tage + tsex

+ tSPO2 + tPaCO2 + tPaO2 + tPaO2/FIO2 +

culture + antibiotic + antibiotic result}

An intelligent agent is constantly monitoring the

respiratory variables and when a set of results occur

the system emits a signal alerting for a possible

infection of the patient. Then the agent waits for the

microbiological cultures results to identify the

bacteria.

If the bacteria verified in the exam it is pseaer,

the agent invokes the function treatments -

responsible to found the possible treatments. This

function compares the biological variables from the

search with the variables from past treatments

(history).

The algorithm returns results where the search

and history variables are closer using priorities in the

calculation of the difference. In this part of the

algorithm it can be returned the treatment data,

considering other factors as the days of treatment

until the antibiotic makes effect and/or the cost of

the treatment.

The variables chosen to compare the results and

their priorities were defined meeting with medical

staff of CHP and they are:

1. SPO2;

2. doi;

3. leuc;

4. PaO2/FIO2 ratio;

5. PaO2;

6. PaCO2;

7. Age;

8. Sex;

This heuristic is represented by the following

algorithm:

Algorithm - Treatment Alternatives

Requires: search variables, history variables

Function Evaluate Infection

Get search variables

3: If SPO2 < 90 and (PaO2 < 75 or

PaCO2 > 75) and PaO2/FIO2 < 250

then

4: For

atient do

5: There is an infection? - Wait for

Cultures Results

6: I

Pseae

is positive Then

7: Function Treatments

8: Else

9: Wait for positive culture

results

10:

11:

12:

End i

End if

End Function

13:

14:

Function Treatments

Get search variables, history of

treatmen

15:

16:

17:

18:

If antibiotic result = success

Treatments = results ordered

by minimum difference between

variables

Else

o successful treatment

19: End i

20: Return Treatments

21: End function

For other bacteria the process of finding the best

treatment is the same. For these types of infections

the algorithm will change heuristic remains the

same. There will be only differences on the variables

used (but the two groups are the same) and on the

first if clause of the heuristic algorithm.

3.3 Results

As above referred, to test this heuristic it was used a

dataset with past treatments applied to inpatients.

Since this study pretend to test the viability of this

concept it was given an input so the system can

search for the results which best suits the input

variables. In future the agent will do this step in real-

time and using online-learning.

UsingDomainKnowledgetoImproveIntelligentDecisionSupportinIntensiveMedicine-AStudyofBacteriological

Infections

585

Table 3 represents the inputs given to the

algorithm – infected patient data.

Table 3: Inputs given to the algorithm.

Variable Description

doi 7

leuc 13

age 65

sex M

SPO2 90

PaCo2 41

PaO2 71

PaO2/FIO2 180

To exemplify the algorithm let’s consider a male

inpatient characterized by: Age = 65 years: Length

of stay=7 days; Leucocytes count = 13x10

/; SPO2 =

90%; PaCO2=41 mm Hg; PaO2=71 mm Hg; and a

PaO2/FIO2=180 mm Hg.

Given these inputs, the algorithm found 23

possible treatments. Table 4 indicates the result from

the search (top 5 results). This table presents the

historical variables values for each suggestion (R1-

R5), the antibiotic administered, the result achieved

(success or failure), the expected days for the

treatment and the treatment cost.

Table 4: Algorithm output.

Variable Results

R1 R2 R3 R4 R5

doi 8 12 7 4 12

leuc 14.5 14.2 13.2 15.2 16.1

age 70 60 80 65 60

sex M M M M F

SPO2 90 91 87 85 91

PaCo2 39.3 39.6 40.0 39.3 39.6

PaO2 71.5 73.3 70.8 57 55

PaO2/FIO2 181 194 196 198 198

Antibiotic Merop Colist Colist Fluco Pipera

Antibiotic Result Suc Suc Suc Suc Suc

Treatment Days 7 11 6 3 11

Cost 210 52 28 8 31

Along with these results the algorithm also gives

information about the support level of the

antibiotics. It is expressed in form of a percentage

and gives an overview of the antibiotics used in the

past and their success. For example, the output about

the antibiotics application for this infection is:

 Meropenem (Merop) – Used successfully 5

times (26.3%); Used Unsuccessfully 14 times

(73.7%)

 Colistin (Colist) – Used successfully 16 times

(52%); Used Unsuccessfully 14 times (48%)

 Fluconazole (Fluco) – Used successfully 10

times (20%); Used Unsuccessfully 40 times

(80%)

 Piperacillin (Pipera) – Used successfully 2

times (12.5%); Used Unsuccessfully 14 times

(87.5%)

4 DISCUSSION

Analysing the obtained results (table 4) it is possible

to observe that the algorithm suggest some treatment

options in this case the usage of colistin,

meropenem, fluconazole and piperacillin. So based

on this information the medical staff can decide on

which antibiotic should be prescribed using as base

the suggestions made by the algorithm.

The approach developed will be improved and

embedded into the INTCare System. The results will

be available anywhere and anytime by whom have

access privileges.

With this algorithm it is possible to observe a list

of treatments that can produce effects in the combat

of pseaer bacteria. By analysing this information the

clinical staff can have an idea about which is the

more adequate antibiotic.

The system proposes some scenarios about

treatments, eliminating at the start past treatments

which not produce any effect or that have very low

taxes of success.

The physicians are always responsible by the

final decision. The heuristic only is used to help

them to take the better decision in the patient best

interest.

5 CONCLUSIONS AND FUTURE

WORK

This study explored an approach using the heuristic

concept to develop a searching algorithm for

combating infections in order to help the doctors

taking better decisions and saving lives. The domain

knowledge combining literature and empirical

knowledge of the medical staff was crossed with

patient and antibiotic data in order to obtain the

rules. Additionally, this approach introduces the cost

factor in the decision criteria.

Preliminary results demonstrate de utility of the

approach and encourage further work in a wider

scope and the integration of the heuristic in INTCare

system in order to complement data mining

predictive models.

ICAART2015-InternationalConferenceonAgentsandArtificialIntelligence

586

The algorithm is represented by a set of rules that

can be easily stored in a knowledge base and

adapted to other infections.

Future work will include the development of an

intelligent agent to search and use:

 other infections and bacteria;

 other patient data to improve the heuristic.

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 and PEst-

OE/EEI/UI0752/2014.

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

Angus, D.C., Linde-Zwirble, W.T., Lidicker, J., Clermont,

G., Carcillo, J., Pinsky, M.R., 2001. Epidemiology of

severe sepsis in the United States: analysis of

incidence, outcome, and associated costs of care. Crit.

Care Med. 29, 1303–1310.

ECDC, 2012. Surveillance of healthcare-associated

infections in Europe 2007. European Centre for

Disease Prevention and Control. Surveillance Report

from ECDC, Stockolm.

Lindbaek, M., 2006. Prescribing antibiotics to patients

with acute cough and otitis media. Br. J. Gen. Pract.

56, 164–165.

Portela, F., Pinto, F., Santos, M.F., 2012. Data mining

predictive models for pervasive intelligent decision

support in intensive care medicine.

Portela, F., Santos, M.F., Machado, J., Abelha, A., Silva,

Á., 2013. Pervasive and Intelligent Decision Support

in Critical Health Care Using Ensembles, in: Bursa,

M., Khuri, S., Renda, M.E. (Eds.), Information

Technology in Bio- and Medical Informatics, Lecture

Notes in Computer Science. Springer Berlin

Heidelberg, pp. 1–16.

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

INTCARE : multi-agent approach for real-time

intelligent decision support in intensive medicine.

SciTePress.

Ray, C. George, 1934- & Ryan, Kenneth J. (Kenneth

James), 1940- & Sherris, John C & ebrary, Inc 2004,

Sherris medical microbiology: an introduction to

infectious diseases, 4th ed, McGraw-Hill, New York.

Romanycia, M.H. J., Pelletier, F. J., 1985. What is a

Heuristic?. Computational Intelligence, 1, 47-58.

Vincent, J.-L., Rello, J., Marshall, J., Silva, E., Anzueto,

A., Martin, C.D., Moreno, R., Lipman, J., Gomersall,

C., Sakr, Y., Reinhart, K., EPIC II Group of

Investigators, 2009. International study of the

prevalence and outcomes of infection in intensive care

units. JAMA 302, 2323–2329.

doi:10.1001/jama.2009.1754.

Vincent, J.-L., Sakr, Y., Sprung, C.L., Ranieri, V.M.,

Reinhart, K., Gerlach, H., Moreno, R., Carlet, J., Le

Gall, J.-R., Payen, D., Sepsis Occurrence in Acutely Ill

Patients Investigators, 2006. Sepsis in European

intensive care units: results of the SOAP study. Crit.

Care Med. 34, 344–353.

UsingDomainKnowledgetoImproveIntelligentDecisionSupportinIntensiveMedicine-AStudyofBacteriological

Infections

587