INTCARE

Multi-agent Approach for Real-time Intelligent Decision Support

in Intensive Medicine

Manuel Filipe Santos, Filipe Portela and Marta Vilas-Boas

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

Keywords: Artificial intelligence, Multi-agent systems, Agent intelligent Decision support systems, Intensive care.

Abstract: For an Intelligent Decision Support System to work in real-time, it is of great value the use of intelligent

agents that cooperate with each other to accomplish their tasks. In a critical environment like an Intensive

Care Unit, doctors should have the right information, at the right time, to better assist their patients. In this

paper we present an architecture for a Multi-Agents System that will support doctors’ decision by in real-

time, guaranteeing that all required clinical data is available and capable of predicting the patients’

condition for the next hour.

1 INTRODUCTION

There is a need to develop an Intelligent Decision

Support System (IDSS) for intensive care to help

doctors decide about the best care to provide their

patients. Such system is expected to reduce medical

errors, to fasten clinical interventions and, overall, to

provide better care by having the required

information at the right place, at the right time. The

INTCare system has this vision and its goals are

towards an IDSS as a Multi-Agents System (MAS).

“Intelligent agents with their properties of

autonomy, reactivity, and proactivity are well suited

for dynamic, ill-structured, and complex

environments (Gao and Xu 2009), such as an

Intensive Care Unit (ICU). This paper presents the

INTCare system as a MAS for intelligent decision

support, an innovative agent-based architecture, and

the prototype system.

Intelligent agent technology is at an appealing

point in its advance. “Intelligence with Interaction”

applications are progressively more being developed

in domains such as meteorology, manufacturing,

war gaming and capability assessment (Tweedale et

al. 2007).

More useful than a system acting intelligently, is

the ability for a system to cooperate in an

environment with other intelligent agents, whether

they are humans or machines (Guerlain et al. 2000).

Furthermore, “it is necessary for a system to be able

to communicate with others, detect and correct

mistakes, and take advantage of others’ abilities.”.

The key is to use distributed artificial intelligence

for decision support (Cortes et al. 2000).

2 BACKGROUND

MAS are computational systems in which several

artificial agents interact to perform predefined tasks

to satisfy some set of goals (Lesser 2003).

INTCare is an IDSS for intensive medicine that

makes use of intelligent agents based on a MAS. For

supporting clinical decision, it makes predictions of

organ failure and outcome and, according to the

predictions, it suggest therapies, treatments and

procedures, by means of DM (Data Mining)

techniques. In order to accomplish its goals, the

system is divided in four subsystems composed by

intelligent agents that cooperate amongst the

system’s modules, as shown in Figure 1.

Subsystems:

 Data acquisition;

 Knowledge Management;

 Inference;

 Interface.

364

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

INTCARE - Multi-agent Approach for Real-time Intelligent Decision Support in Intensive Medicine .

DOI: 10.5220/0003182603640369

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 364-369

ISBN: 978-989-8425-40-9

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: INTCare system and its four subsystems.

The data required for the DM models is gathered

in the ICU and relates to the patients’ clinical

condition (monitored data, nursing records, lab

results) and derived variables like Sequential Organ

Failure Assessment (SOFA) scores, critical events

(Silva et al. 2008), ratios (Vilas-Boas 2010) and

Modified Early Warning Scores (MEWS) (Subbe et

al. 2001). These data origins in distributed

heterogeneous data sources and, to guarantee the

system’s viability in real-time, it relies on intelligent

agents. In this paper, we focus on latest

developments of the data acquisition subsystem in

order to achieve INTCare’s goals. They go beyond

retrieving, collecting and integrating relevant

information; they encompass discovering previously

unknown, implicit and valuable knowledge (Klusch

et al. 2003). Even though the previous system had a

data acquisition subsystem, it lacked some important

features and functionalities, mainly related to the

information it was able to acquire. This subsystem

was able to gather and store data from the vital signs

monitoring. However, with our advances in the

research related to the prediction of organ failure,

there is the need for more data other than the

previously gathered, mainly data from lab results

and drugs systems. Also, previously, data was

collected and post processed, which was not suited

for a real-time system.

The system is currently being tested in the ICU

of Hospital Geral de Santo António (HGSA), Porto,

Portugal.

3 DATA ACQUISITION SYSTEM

- REDESIGN

With the new requirements of INTCare, i.e. data

collected in real-time and available online, it was

necessary to redesign the data acquisition system

and its architecture. The new model allows all data

to be collected and stored in electronic mode and the

extinction of paper-based information throughout the

ICU. The existing agents were preserved and added

new tasks.

Figure 2: Data acquisition subsystem.

The tables contained in the database are defined

in terms of their attributes, as follows:

HL7_HDR  {MSG DATA}

HL7_MSG  {Vital Signs)

UCI_PATIENT_FEE  {PaO

, FiO

, Glasgow,

Urine Output, Amines,

Patient Outcome, Drugs, …}

UCI_PATIENT_LR  {Creatinine, Billirubin,

Blood Platelets, PaO

, FiO

SOFA, …}

UCI_DATABAE {HL7_MSG,UC

I_PATIENT_FEE,

UCI_PATIENT_LR}

Figure 2 presents the link to other subsystems,

which is a consequence of the work of Data

Acquisition System.

4 INTCARE SYSTEM’S AGENTS

The tasks of the intelligent agents support some

system modules: Data Acquisition, Management of

Knowledge, Inference and Interface. The flexibility

and effectiveness of such systems depend on the

agents and the interactions between them.

4.1 Data Acquisition Agents

These agents will be in charge of the tasks

associated to the data acquisition.

INTCARE - Multi-agent Approach for Real-time Intelligent Decision Support in Intensive Medicine

365

Gateway (a

gat

)

is responsible for capturing the vital

signs data from bedside monitors and operates in

real-time. These data are packed into HL7 messages

and sent to the Vital Signs Acquisition Agent. It

collects information, in average, once per minute

and restarts at each hour to ensure that no

communications failures compromise the system;

Vital Signs Acquisition (a

vsa

)

was reformulated. It

is an AIDA process that parses the HL7 messages in

real-time, extracts information blocks, splits the

information and stores them in the database tables.

The table HL7_Message contains the values and

results from vital signs monitoring (id_category,

value, patient_id, msg_date).

ENR Agent (a

enr

) is new and was created with the

Electronic Nursing Record (ENR) (Portela et al.

2010, M.F. Santos et al. 2009b) and replaces the

Clinical Data Entry Agent. It was developed to allow

some information in electronic mode.

This agent works in two forms: automatic (can

save the data in minutes) and manual, when it’s the

user who saves the information. It is responsible to

capture the clinical data from the medical and

nursing staff (Gago et al. 2006). When it receives the

data, it splits the information, clears the information

displayed in the table for that patient in that day and

stores the new data.

LR (a

) is new and it’s an evolution of Clinical

Analysis Data Entry Agent that was responsible for

capturing the clinical data from the lab results.

This agent works in association with the ENR

because it only gets information from a certain

patient when it is required by the ENR. The process

is easy and smooth; The ENR agent requests to (a

every five minutes, new results of lab results. This

agent verifies if there are new results from a patient

and, if so, it stores them in UCI_PATIENT_LR and

sends the new information to ENR.

All information collected and stored by data is

available by AIDA (Abelha et al. 2003).

AIDA (a

ada

) is an agency to archive and to

disseminate medical exams and results, implemented

at the hospital. It supplies the lab results and nursing

records through the clinical analysis data entry agent

and clinical data entry agents (Abelha et al. 2004).

Pre-processing (a

) agent is responsible for the

correct linking of all the values in order to create a

valid medical record for the patient (Gago et al.

2006). It is in charge of solving some data

acquisition problems (M.F. Santos et al. 2009a).

Before data is consolidated in the data

warehouse, the agent verifies the data in order to

remove null values and correct the values that are

out of range (M. F. Santos et al. 2009). It proceeds

with the copy of the values received from the three

data sources: bedside monitors, electronic nursing

records and lab results that were stored in the tables.

At the moment, we are studying another method

to optimize results using the MEWS.

4.2 Knowledge Management Agents

This subsystem is composed by three agents: Data

Mining, Performance and Ensemble.

Data Mining (a

) agent is responsible for the

retrieval of the relevant data in order to make

possible the application of AI algorithms for the

predictions and to train new models, whenever

requested by the Performance Agent (apf). After the

training, the models are stored in the the Knowledge

Base, applying the Predictive Markup Model

Language (PMML), which is a XML based model

description language (Guazzelli et al. 2009) (Gago et

al. 2006).

This agent converts the data stored in the data

warehouse in Knowledge, creating models in real-

time using online data.

Performance (a

) agent is responsible for

continually consulting the data warehouse to collect

statistics, as a base to calculate a set of assessment

parameters maintained in the Performance Database

(Gago et al. 2006). It analyzes the new data that was

stored in data warehouse and verifies the

performance of the prevision models through DM.

If the collected statistics show that the performance

has fallen below a predefined parameter, a new

model is requested to replace the poor performing

one (Gago et al. 2006).

Ensemble (a

ens

) agent was created to enhance

predictive performance by combining several

models in order to produce models with better

results (Dietterich 2000); It is possible to combine

techniques like decision trees and artificial neuronal

networks.

4.3 Inference Agents

The inference subsystem didn’t have any

adjustments and its agents are the same presented

with the INTCare System (Gago et al. 2006).

Prediction (apd) agent answers user questions by

applying the adequate models contained in the

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

366

Knowledge Base to the data stored in the data

warehouse (Gago et al. 2006);

Data Retrieval (adr) agent’s task is to retrieve, from

the data warehouse, the information requested by the

interface agent (Gago et al. 2006);

Scenario Evaluation (asc) agent makes possible for

doctors to create and evaluate what-if scenarios.

After receiving the data from the interface agent, this

agent requests a forecast from the prediction agent,

the scenario is then stored in the Scenarios Database.

and the result is sent back to the interface agent

(Gago et al. 2006).

4.4 Interface Agents

The Interface subsystem only has one agent, the

Interface Agent. This subsystem suffered some

alterations due the changes presented before. At the

moment we have the web-based ENR that collects

data from various data sources. ENR is also a

consulting platform for patients’ clinical data and,

furthermore, the prediction models created by DM

can also be consulted.

Interface (aint) agent makes possible web-based

interaction with the system by providing an easy

way for doctors to request prognostics and evaluate

scenarios (Gago et al. 2006). For the past days, the

actual SOFA score is shown, together with the value

predicted for the next day. It is also possible to

consult the confidence levels associated with the

prediction values (Gago et al. 2006). At the moment,

we can predict the results for the next hour (Vilas-

Boas 2010) and a holistic view of the patient to the

next hour will be supported by presenting, in the

same chart, the information regarding the organic

systems. Also, the values predicted for the next hour

are shown, providing information that may suggest

the need for immediate action.

5 MULTI-AGENT SYSTEM

INTCare is a MAS because it uses a conjunction of

several agents for preparation of data for prediction.

MAS consists of a set of agents that communicate

with each other and work together towards common

goals, with a degree of reactivity and / or reasoning

(Wooldridge, 1999).

5.1 First Multi-agent Process

All agents presented collect data in real-time and

online mode. They communicate among themselves

to obtain better results. The a

gat

agent processes the

monitored data. When the gateway receives the vital

signs from the monitors, it sends a HL7 message

(M1) to the vital signs acquisition agent and it parses

the information. Next, there is an example of a HL7

message:

MSH|^~\&|DHV |h2|h3|h4|||ORU^R01|h1|P|2.3.1

PID|1||d1||d2

PV1|1|U|v1

OBR|1|||DHV|||r1|

OBX|x2|NM|x3^x4^^^x5||x6|x7|||||R||||x1^ v1||

Table 1 explains the variables in the exchange of

messages between the agents a

gat

and a

vsa

Table 1: HL7 message variables (M. F. Santos, et al.,

2009).

h1 Version ID x1 Producer's ID

h2 Sending Facility x2 Value Type

Receiving

Application

Observation

Identifier (cod)

Receiving

Facility

Observation

Identifier (cod2)

Patient ID

(Internal ID)

Observation

Identifier (descp)

d2 Patient Name x6

Observation

Value

Assigned Patient

Location

Units

Observation

Date/Time

The information received by a

gat

is split by a

vsa

the form:

HL7_HDR = {h1, h2, h3, h4, d1, v1, x1 }

HL7_MSG = {v1,r1,x2,x3,x4,x5,x6,x7,x1^v1}

5.2 Second Multi-agent Process

The ENR agent communicates with two other agents

vsa

and

) and all tasks are triggered by the ENR.

The ENR comprises of two forms to register data

- manual and automatic. When the ENR requires

information to fill its fields, it sends requests to the

ENR Agent with the alert that data it required.

The a

enr

analyzes the request and, according the

needs, (M2) sends information to the other agents

(M3, M5).

The Vital Sign Agent works in real-time and has

new data all the time. In this case, it sends the new

data (M4) to a

enr

and it is sent to the ENR. If the

ENR need lab results, the ENR Agent sends the

Patient ID (PID) (M5) to Lab Results Agent and it

INTCARE - Multi-agent Approach for Real-time Intelligent Decision Support in Intensive Medicine

367

will verify if some new results are available in Lab

Results Server.

Finally, the a

stores data in the DB table and

sends the newest data to the ENR Agent (M6)

which, in turn, sends to the ENR application.

5.3 Other Multi-agent Process

The knowledge management, inference and

interface subsystems needn’t modifications at the

moment, for they are working correctly and in

accordance to the MAS and INTCare’s goals. Their

communication is visible in the arrows presented in

the agents of each subsystem in figure 1.

 Performance agents perform some actions to

Data Mining Agent and Ensemble Agent;

 Assemble Agents perform some actions to data

Mining Agents;

 Scenario Evaluation Agent works together with

Prediction Agent and Interface Agent;

 Prediction Agent works with interface agent and

Scenario Evaluation Agent;

 Data Retrieval Agent works with Interface Agent

and vice-versa;

 Interface Agent works with Scenario Evaluation,

Prediction Agent and Data Retrieval Agent;

6 DISCUSSION

Table 2: Impact of the MAS system in intensive care.

Characteristics Impact

Robustness

The medical staff can work with the

guarantee that the information will

never lost.

Fault Tolerance

Lack of system failures in critical

environments and reduction of medical

error.

Scalability

Possibility to work with large

quantities of clinical information.

Speed up a system's

operation

less time spent with informatics issues.

Efficiency

Systems always running in real-time

and efficiently doing what users

require.

Interconnection and

interoperation of

multiple existing

legacy systems

Medical Staff can have information

about patients when required,

independently of the systems in which

they that are stored.

Table 2 summarizes a set of characteristics of the

MAS and their impact on a critical environment like

ICU, comparing to a single agent or centralized

approach.

By distributing computational resources and

capabilities across a network of interconnected

agents, performance bottlenecks or critical failures,

are not a problem (Gruer and Hilaire 2003).

Overall, with the MAS there is great gain in

system’s performance, efficiency, reliability,

scalability, robustness, maintainability,

responsiveness, flexibility, and reuse (Chira 2007).

These are a critical aspects in intensive care

where, most of the time, patients’ lives are at risk

and all valuable information should be available and

its correctness and ease of interpretation must be

guaranteed. Moreover, the MAS will also provide

the means for prediction and decision models in

real-time, with an hourly granularity.

In our approach, each of the intelligent agents is

able to deal with specific tasks that can be grouped

in data acquisition, knowledge management,

inference and interface in an independent way, so

that the system has a transparent behaviour for the

applications, as well as for the users. For a real-time

system like INTCare, a MAS presents advantages in

processing speed-up, automatic data validation,

access to data and availability of information

electronically.

All modifications were tested in the real

environment of the ICU, guaranteeing that they will

suit correctly the requirements and identifying

possible new requirements and/or refinements.

7 CONCLUSIONS

AND FURTHER WORK

For accomplishing INTCare’s goals, some redesign

regarding the agents was presented. Data acquisition

subsystems and its agents were reformulated and

new agents were created.

Agents can perform some previously manual

action in automatic mode, without human

intervention. This feature is of great importance in a

real-time IDSS. To achieve so, the Data Acquisition

Architecture was redesigned to allow data

acquisition in real time and online mode.

Regarding the interface, we are studying two

solutions for presenting the information originated

by the DM models – partially in the ENR and

wholly it INTCare platform.

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

368

ACKNOWLEDGEMENTS

The authors would like to thank FCT (Foundation of

Science and Technology, Portugal) for the financial

support through the contract PTDC/EIA/72819/

2006.

REFERENCES

Abelha, A., Machado, J., Alves, V. and Neves, J. (2004)

'Health data management in the medical arena',

WSEAS-Transactions-on-Computers.

Abelha, A., Machado, M., Santos, M., Sollari, A., Rua, F.,

Paiva, M. and Neves, J. (2003) 'Agency for Archive,

Integration and Diffusion of Medical Information',

Proceeding of AIA.

Balter, J., Labarre-Vila, A., Ziébelin, D. and Garbay, C.

(2002) 'A knowledge-driven agent-centred framework

for data mining in EMG', Comptes Rendus Biologies,

325(4), 375-382.

Chira, C. (2007) 'A Multi-Agent Approach to Distributed

Computing'.

Cortes, U., Sanchez-Marre, M., Ceccaroni, L., R-Roda, I.

and Poch, M. (2000) 'Artificial intelligence and

environmental decision support systems', Applied

Intelligence, 13(1), 77-91.

Dietterich, T. (2000) 'Ensemble methods in machine

learning', Multiple classifier systems, 1-15.

Gago, P., Santos, M. F., Silva, Á., Cortez, P., Neves, J.

and Gomes, L. (2006) 'INTCare: a knowledge

discovery based intelligent decision support system for

intensive care medicine', Journal of Decision Systems.

Gao, S. and Xu, D. (2009) 'Conceptual modeling and

development of an intelligent agent-assisted decision

support system for anti-money laundering', Expert

Systems with Applications, 36(2), 1493-1504.

Gruer, P. and Hilaire, V. (2003) A multi-agent approach to

modeling and simulation of transport on demand

problem, translated by Springer, 119.

Guazzelli, A., Zeller, M., Lin, W. C. and Williams, G.

(2009) 'Pmml: An open standard for sharing models',

The R Journal, 1(1), 60-65.

Guerlain, S., Brown, D. E. and Mastrangelo, C. (2000)

Intelligent decision support systems, translated by

1934-1938.

Klusch, M., Lodi, S. and Moro, G. (2003) 'Agent-based

distributed data mining: The KDEC scheme',

Intelligent information agents, 104-122.

Lesser, V. R. (2003) Multi-agent systems, translated by

John Wiley and Sons Ltd., 1196.

Portela, F., Vilas-Boas, M., Santos, M. F. and Fernando,

R. (2010) 'Improvements in data quality for decision

support in Intensive Care', in eHealth 2010 - 3rd

International ICST Conference on Electronic

Healthcare for the 21st century, Casablanca, Morroco,

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

Abelha, A. and Neves, J. (2009a) 'Information

Architecture for Intelligent Decision Support in

Intensive Medicine', 8th WSEAS International

Conference on APPLIED COMPUTER &

APPLIED COMPUTATIONAL SCIENCE (ACACOS

'09), 8(5), 810-819.

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

Abelha, A., Neves, J., Silva, A. and Rua, F. (2009)

'Information Modeling for Real-Time Decision

Support in Intensive Medicine' in Chen, S. Y. and Li,

Q., eds., Proceedings of the 8th Wseas International

Conference on Applied Computer and Applied

Computational Science - Applied Computer and

Applied Computational Science, Athens: World

Scientific and Engineering Acad and Soc, 360-365.

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

Abelha, A., Neves, J., Silva, A., Rua, F., Salazar, M.,

Quintas, C. and Cabral, A. (2009b) 'Intelligent

Decision Support in Intensive Care Units - Nursing

Information Requirements' in Springer, ed. WSEAS

Transactions on INFORMATICS.

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

J. (2008) 'Rating organ failure via adverse events using

data mining in the intensive care unit', Arificial

Intelligence in Medicine, 43(3), 179-193.

Subbe, C. P., Kruger, M., Rutherford, P. and Gemmel, L.

(2001) 'Validation of a modified Early Warning Score

in medical admissions', QJm, 94(10), 521.

Tweedale, J., Ichalkaranje, N., Sioutis, C., Jarvis, B.,

Consoli, A. and Phillips-Wren, G. (2007) 'Innovations

in multi-agent systems', Journal of Network and

Computer Applications, 30(3), 1089-1115.

Vilas-Boas, M., Santos, MF.,Portela, Filipe., Silva,

Álvaro., Rua, Fernando. (2010) Hourly prediction of

organ failure and outcome in intensive care based on

data mining techniques, translated by Funchal,

Madeira.

INTCARE - Multi-agent Approach for Real-time Intelligent Decision Support in Intensive Medicine

369