HOW TO BUILD A MODERN PATIENT CARE APPLICATION

Dieter Gawlick, Adel Ghoneimy and Zhen Hua Liu

Oracle Corporation, 500 Oracle Parkway, Redwood City, CA 94065, U.S.A.

Keywords: Declarative programming, Data management, Knowledge management, Tagging, Meta-queries.

Abstract: Patient care is typically supported by several incompatible applications in terms of their data models and

semantics, ranging from support for surgery, to ICUs, to standard health care. These programs are focused

on specific situation and give doctors a very limited view at patient information. This paper argues that the

IT technology has evolved to a point that it is now possible to develop a generic patient care application that

manages all patient data for all situations. Furthermore, this application provides a framework for capturing

medical knowledge. With this knowledge the application is able to extract medically relevant information

from data, even if the extraction is outside of the medical expertise of a doctor or if the extraction is outside

of the capability of the human brain. Medical knowledge can be customized by domain, personal views, and

patient condition. Additional benefits are: alerting the right doctor of time critical situations, identifying

relevant information from a (new) patient’s history, and timely sharing of (new and codified) medical

knowledge and experience across the medical community. The ideas are based on prototype work between

the University of Utah, Health Science Center (UUHSC), the University of Coimbra, and Oracle (Guerra et

al., 2009).

1 INTRODUCTION

Today’s patient care applications are focused on

managing patient data using only the most basic

functionality of database systems. Operational tasks,

such as monitoring patient conditions, alerting

doctors for critical patient conditions, classifying

and extracting information from patient data and

developing and supervising treatment plans, are

considered to be beyond the scope of DBMS.

Therefore, a significant amount of procedural code

is written leading to limited functionality, lack of

customization, and extremely high development and

maintenance costs. Long development cycles are

another side effect.

In this paper we show that modern database

technology has matured to a point that even

operational tasks can be handled by databases based

on declarative specifications.

Databases can capture a wide variety of data

since they support variety of data models; SQL for

the management of highly structured data, XML for

the management of semi-structured data (Liu et al.,

2005), and RDF/OWL for formal semantics and

reasoning (Das, S.

et al., 1004); DICOM can be

used to manage medical images. Ad-hoc access to

single records is complemented by a rich support for

analytics, all the way to the support for the

development of (non-hypothesis) driven predictive

models and machine learning. Databases are also

able to automatically manage the history of patients

providing easy and fast access while minimizing

storage consumption; updated and deleted records

are simply remembered.

Databases have a variety of ways to extract

information from raw data in near real time.

Examples are triggers, real-time log-miners,

registered queries, online scoring, and RDF/OWL

rules (Oracle, 2010b). Any of these technologies can

be seen as a way to capture domain knowledge and

apply them to the data; i.e., databases not only

manage data but also manage knowledge and apply

this knowledge immediately to incoming data (

Motro

and Yuan, 1999). Extracted information can be stored

in the database as well, however, extracted

information will typically be represented by an

abstract vocabulary that most likely consists of a set

of tags; e.g., the vitals of a patient are critical. The

association of tags to data is often the core of

applying domain knowledge to captured data. Time

critical information will trigger alerts. New or

improved knowledge will trigger a re-evaluation of

existing data and trigger alerts as well.

427

Gawlick D., Ghoneimy A. and Liu Z..

HOW TO BUILD A MODERN PATIENT CARE APPLICATION.

DOI: 10.5220/0003298404270432

In Proceedings of the International Conference on Health Informatics (HEALTHINF-2011), pages 427-432

ISBN: 978-989-8425-34-8

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

Knowledge is widely shared among experts;

however, it is important that knowledge can be

tailored to the view of sub-domain specialities and

individuals. This can be done by integration of

knowledge management and consumer management.

The rest of the paper is organized as followings.

Section 2 describes the motivating example of

patient care. Section 3 describes important aspects of

advanced data management, knowledge

management, and consumer management, section 4

shows how these technologies are used

synergistically, section 5 compare this approach to

others and section 6 draws conclusions.

2 PATIENT CARE

The capturing of patient information with EMRs is

on its way to become the baseline for using IT

technology for patient care. Ideally, the medical

personnel and patients should have unobstructed

access to any information they are entitled to see;

e.g., they should have access to individual records –

present and past as well as support for higher level

functions such as trends and scoring against any

number of conditions.

This management of and access to patient

records has to be complemented by the management

of the medical knowledge, i.e.., the system has to

support the medical personal and patients in the

interpretation of data. Here are two examples: the

application should be able to determine if a specific

value, or a set of values indicate a risk, i.e., create if

necessary a tag indicating a serious or critical

condition. The application should also to able to

identify those models – diagnosis’s - for which a

patient scores high. This application of the medical

knowledge has to be done whenever new data

becomes available. If there is any urgency in the

computed information, the right members of the

medical personnel have to be alerted.

The interpretation of data is obviously a very

complex process; it depends not only on patient data

but also on the diagnosis, the speciality, as well as

the preferences of a doctor. Consequently, the

knowledge management has to capture shared and

generally accepted knowledge and also provide the

freedom for individual points of view.

The knowledge that is used to support an

individual doctor can be considered as a profile of

that doctor. Profiles should be used to significantly

improve query processing. Instead of looking just at

data, doctors should be able ask for important

information that are (hidden) in the patient data; as

of the time of the query or in the past. This would

allow doctors to find important information fast,

even in a large amount of data. Missing data (lab-

test) should also be identified.

Since the medical knowledge evolves extremely

fast; the application must not only be able to adapt

easily to new and improved knowledge but also help

doctors to improve their knowledge by referencing

background information about the applied

knowledge.

All information derived by the knowledge

management system has to be documented for

proper auditing, tracking, and provenance; data,

knowledge, and derived information have to be lined

up in time.

Such an application could provide a better

foundation for evidence-based medicine.

Today, knowledge is shared by writing papers.

There are just too many papers to read and the

information is often not precise enough. Papers have

to be complemented by a representation in machine

useable form. This representation – with the proper

precaution and adjustment – should be easily

importable into a knowledge base.

Models – applied to a large number of patients -

should be used to provide insight in the selection of

the best ‘standard of care’ even taking patients’

objectives into account.

3 THE TECHNOLOGIES

The development of new medical applications needs

to leverage these three core technologies: data

management, knowledge management, and

consumer management.

3.1 Data Management

3.1.1 Data Models and Extensibility

Medical applications require the management of

many types of data, far beyond classical SQL92,

which is used by most applications and only

supports simple structured relational data. Modern

DBMSs have evolved to a point that they can store,

manage, and query all kinds of data in one system,

including DICOM data that are specifically designed

for medical applications. Table 1 shows a few types

of data that healthcare applications need with the

corresponding support in modern DBMSs.

Storing data using these technologies without

any intervening application logic is the preferred

way of managing data. It provides an unobstructed

HEALTHINF 2011 - International Conference on Health Informatics

428

Table 1: Use of data types.

Data Model with Query

language supported by modern

DBMS

Data instance examples in

Healthcare Application

SQL92 relational data with

SQL

Patient id, name, birth-date

etc

XML with XQuery &

SQL/XML

Doctor notes, comments with

XML tagging

RDF/OWL with SPARQL Knowledge management

Domain specific data with

object SQL and extensibility

DICOM patient data

view at data.These technologies are typically built

on top of the database extensibility framework

(Stonebraker et al., 1998).

3.1.2 Data Mining and (Online) Scoring

Information can be extracted from records by

defining expressions (rules); however, what can be

done to extract complex information, i.e., I need to

assess the likelihood of one or more patients having

a cardiac arrest within the next 24 hours? It is close

to impossible to formulate the right expressions

(rules) manually; data mining can be leveraged for

the development of (non-hypothesis) models as well

as scoring these models based on patient

information. Modern DBMS can support data

mining efficiently (Milenova et al., 2005).

3.1.3 History Management

Capturing the history of specific patient data with

program logic is a very resource intensive in terms

of development effort (typically > 50% of this

effort) as well as storage and access cost.

Transactional temporal databases, i.e., databases that

keep previous and deleted versions of records and

make them easily accessible are a much better

approach. Additional benefits can be gained from

RDBMS with full ILM (Information Life Cycle)

support, including compression, access optimization,

and more. Data history support provides the base for

auditing, tracking, and provenance. A good example

for temporal support is Total Recall (TR) in the

Oracle database (Venkata et al., 2008).

3.1.4 Event Processing

Databases have significant support for event

processing. Trigger and (online) log mining are

widely available technologies (Chandy and Gawlick,

2007). A more recently introduced capability, is to

if there is a change in an object or a result set

(Oracle, 2010a). While registered queries allow

extracting important information from databases in a

timely fashion. Materialized view capability can

handle more computational intensive extraction in

periodical fashion.

3.1.5 Tagging and Grouping

Annotating information with tags is a flexible way

of classifying and grouping information; e.g.,

normal, guarded, serious, and critical is a

meaningful tag set to classify a wide range of patient

data. Using this tag set one can ask for any data

values that are classified as serious. Another

important tag set is the diagnosis. These tags could

be complemented by information about uncertainty,

relevant data/tests to reduce any uncertainty, and

other information.

Tags reduce significantly the complexity and the

number of queries. This is due to the fact that a

valuable part of the domain knowledge is captured

in the selection of the tag attributes, the tag sets, and

the calculation of tag values. While there is a wide

consensus within a domain about these selections

and calculations, it is important that customization

for special areas, groups, and individual views is

well supported and easy to use.

Tagged data can be modelled using XML data

model and querying of tagged data can leverage

XQuery that is designed to query XML (Liu et al.,

2005) while RDF/OWL technology can be leveraged

to support customization.

3.2 Knowledge Management

The success of modern patient care applications

requires effective management of domain

knowledge. This section explains the essence of

(operational) knowledge lifecycle management, and

how such knowledge can be applied to data and

produce relevant timely information.

3.2.1 Knowledge Lifecycle

Knowledge Capture: Before knowledge is captured

in the form of logical assertions, an ontological

model (T-Box) must be specified, to provide the

logical foundation for the assertions to be captured.

Such model may initially be very simple. As Subject

Matter Experts (SMEs) gain more knowledge about

better ways of interpreting the facts in the database,

the model can be evolved to support such learning.

Support for personalized ontological model and

assertions are essential for lowering the barrier to

capturing knowledge. Community or institutional

knowledge is eventually derived from the

aggregated personalized knowledge. Also, allowing

HOW TO BUILD A MODERN PATIENT CARE APPLICATION

429

SMEs to record their subjective belief in a

generalized manner without restrictions, lower the

barrier to capture such belief. Such generalized

knowledge can be eventually specialized at a later

time by subjecting them to deductive rules and

machine learning techniques.

Knowledge Validation: Validating captured

knowledge and assigning pedigree to its models and

assertions is essential to the trustworthiness of such

knowledge. It can be achieved through ontology

validation heuristics, computing statistical

consensus, or social network analysis. Increasing the

trustworthiness and consistency of knowledge

content tend to increase the knowledge base

resistance to invalid or incorrect assertions and

models.

Knowledge Retrieval: Knowledge retrieval is

essential for sharing, transferring, and learning the

knowledge of a community of practice or an SME.

Database embedded RDF/OWL technology

simplified, to great extent, the browsing, querying,

and searching of such knowledge (Das, S. et al.,

1004), (Oracle 2010b). While an NLP based search

interface is not within the scope of this paper, it is

essential to effective knowledge search.

Knowledge Maintenance: Knowledge

maintenance is essential for the evolution of SMEs

and community understanding and interpretation of

the facts. The ability to specify defined classes and

infer membership and subsumption of such defined

classes is extremely important to evolving the

personal and institutional knowledge. Such

capabilities facilitate the refinement of the

knowledge and enable the encoding of deeper

understanding of the facts on ongoing basis.

3.2.2 Knowledge Application

In our approach, knowledge is maintained in two

representations; a formal model that is platform

independent and an executable model that is

platform specific. Both models are required to be

declaratively specified making it possible to

maintain consistency between the two models,

should one of them changes. As SMEs and

Knowledge curator maintain and evolve the

knowledge, a system would be able to reflect such

evolution on the platform specific model using

automatic transformation rules.

Similarly, changes to the platform specific model

by IT personnel and data mining specialist can be

reflected on the platform independent model. This

will also allow us to test the validity and consistency

of such change.

In the following section we will show that the

current database technology is able to encode

operational knowledge effectively and to process

data transactionally in real-time using deductive

reasoning, event processing pattern recognition, and

meta-queries.

Deductive Reasoning: Deductive database

research recognized the need for applying deductive

reasoning to deduce actionable information from the

database. It facilitated the application of logic and

inference to relational data (Tsur, 1991)

(Ramamohanaro, K. and Harland J. 1994). Modern

databases (Oracle, 2010b) can support the

transactional application of deductive reasoning in

real-time. Semantic Assisted Queries is one

example. Such capabilities have enabled the creation

of new breed of applications capable of applying

domain knowledge to the facts in the database to

extract humanly consumable information in real-

time (Universitaet Bonn, 2010).

Event Processing: Timely awareness of critical

incidents – alerting - is very important in the medical

and almost any other domain. Alerting requires users

to specify which information is important to

him/her. Database triggers represent the first

generation of active data supporting explicit event

processing paradigm (Chandy and Gawlick, 2007).

Though trigger support is effective, it is still too low

level and is not scalable with large number of

triggers. Instead, second generation event processing

in active data management enables user to register

large number of patterns and efficiently evaluate

them in a scalable manner via indexing over

conditional Expression. Patterns of interests can be

defined using conditional expressions

A data type that is of special interest for patterns is

the expression type (Yalamanchi et al., 2003

Expressions can be stored in tables with a column of

type expression. Given input data, one can find all

the stored expressions that are evaluated to be true

for these data. The scalable is achieved through a

specific index for expressions. Expression filter are

very effective in supporting personalized multi-level

pattern recognition preferences. Event Processing

facilitates the encoding of rules knowledge in the

form of Event-Condition-Action.

Continuous Query Notification (CQN) on the

other hand continuously evaluates the state of an

object or a query and alert users and application in

the event of occurrence of such change (Oracle,

2010a). Materialized view can be used to derive new

knowledge given new patterns are found.

Meta-Queries: As a large body of (medical)

knowledge is captured and encoded in declarative

HEALTHINF 2011 - International Conference on Health Informatics

430

rules and queries, it is important to apply such

knowledge to (newly arriving) patients with a large

medical history. Therefore, there is a need for

queries that reference the knowledge base as well as

the fact base to extract important information. We

call these queries meta-queries. The result of these

queries is a set of records that indicate which

elements of a knowledge base identified relevant

queries and what was found; e.g., a meta-query may

identify a set of diagnosis, their time, and even

associate probability.

It is not feasible to apply all rules to all patient

records due to the length of time it takes to apply all

known rules and queries to patient records. One way

to find such relevant queries is to treat the rules and

queries as an abstract data type, build an efficient

index, and identify relevant queries by using patient

data to traverse this index.

Meta-queries provide the same results as event

processing if the same queries and data are used.

The only difference is the time results become

available.

3.3 Consumer Management

An application that supports data and knowledge

management can deduce information from raw data

as soon as new information becomes available. This

interpretation of data, however, often depends on the

personal view and preferences of individual doctors

– the consumers of the information. Therefore it has

to be known which doctors are responsible for which

patients in order to apply the right profiles (of

knowledge) to incoming data. If the deduced

information is classified as being urgent doctors (or

other medical personal) have to be alerted

immediately. Obviously, the deduced information as

well as the urgency can be different for different

doctors –allowing doctors in a team to share each

other’s view may be beneficial.

This function requires up-to-date data about the

consumers (medical personnel), their assignment,

their availability, how to reach them, as well as their

(evolving) profiles; i.e., the status of all consumers

and their organizational context must be known at

any point in time.

4 PUTTING IT ALL TOGETHER

Applications depend on four main components;

knowledge, data, information, and consumers.

Knowledge Base contains the structure and

semantics of the data and all the codified domain

knowledge. The codified knowledge can be

represented in the form of rules, expression (if

expression then …), models or any other means

useful to extract information from data. The

knowledge base is highly customizable, it contains

the codified knowledge of the entire medical

community customized by views and interest of

doctors working in specific domains, as well as

customized by the view of individual doctors, and

even the way doctors looking at specific patients.

Data contains all the measurements and

observations that are captured by sensors and the

medical personal; here we will record the vitals, the

blood chemistry, and whatever else has been

observed. Data supports a wide range of data types,

history management, life cycle management as well

as all essential operational characteristics, such as

performance, scalability, availability, fault/disaster

tolerance, and security.

Data Information

KnowledgeManagement

Consumers

Figure 1: Application components.

Information contains insights that have been

derived from applying the operational knowledge;

here we will find a note if the heart beat is perceived

as critical or if a patient has a high risk of a cardiac

arrest within 24 hours. The underlying functional

and operation support is at the same level as for data.

Consumer contains details about the medical

personnel, such as who is on shift, who is available,

and how to reach a doctor or nurse. Most

importantly there is data on the customization of the

operational knowledge.

Raw data will be captured using OLTP or

protocols; export is possible as well.

Whenever new data are captured, they and the

previously captured data will be analyzed using the

operational knowledge, if anything of importance is

found, one or more entries in information base will

be made; e.g., data will be transformed into

information. It should be noted that information

entries could be created even if there is no change in

data. For example, the blood pressure has remained

high for an extended period. Doctors can add

informational notes as well.

Entries in the information base may be tagged as

time critical; so as to alert attending doctors/nurses.

In addition to being alerted, the medical

personnel can access data and information entries at

HOW TO BUILD A MODERN PATIENT CARE APPLICATION

431

any point in time. If a doctor gets an alert he/she

may be interested in the background leading to this

alert, the system will be able, not only to show

which data triggered the alert, but also which entries

in the knowledge base were used. This is especially

helpful if an alert is triggered by a knowledge base

entry representing medical knowledge a doctor is not

familiar with.

Obviously, doctors can ask questions referring to

information such as what is important to know about

this patient; rather than going through 1000’s of

EMRs.

5 COMPARISON

It may be interesting to compare our approach to

OLTP and (stream based) event processing. OLTP

only manages data; there is neither a knowledge

base, nor information, nor consumer management.

Event processing supports the data component only

in a very specific way; entries in data are transient

and ordered by a time stamp, there is no query

support and the focus for operational characteristics

is only on performance. Knowledge management is

implicitly represented as a set of rules or continuous

queries, results from queries/rules are thrown over

the wall; there is no information and consumer

management. Our approach is integrated and it is

based on deductive database concepts enhanced by

tagging, grouping, and consumer management. This

direction is aligned with the work in (Universitaet

Bonn, 2010).

6 CONCLUSIONS

The ISCU prototype has shown that it is possible to

develop a patient care application that can be

universally used in ICU environments and other

patient care departments in hospitals, institutions,

out-patient care, as well as long term home care.

This application manages data, knowledge,

information, and consumers. It captures data,

deduces information from these data based on

personalized knowledge, alerts if urgent actions are

required, and provides unobstructed access to data,

information, and knowledge.

Our approach of integrating data, information,

knowledge, consumer management enables medical

applications to be developed in declarative manner

so that domain experts, such as medical staff, can

share knowledge and experience – easily, flexibly

with little delay. Such application can evolve

without dependency on IT personal and adjust to

advances in IT technology.

ACKNOWLEDGEMENTS

We like to thank Ute Gawlick (MD/PhD) for

ensuring that the SICU prototype reflects the needs

and thinking of the medical community, Diogo

Guerra for developing a complex prototype in only 4

months, and Pablo Tamaya for developing a non-

hypothesis driven predictive model for cardiac

arrest. Their work is the base for this paper.

REFERENCES

Chandy, K., Gawlick D., 2007. Event processing using

database technology. SIGMOD Conference 2007:

1169-1170.

Das S., Chong E., Eadon G., Srinivasan, J., 2004.

Supporting Ontology-Based Semantic matching in

RDBMS. VLDB 2004: 1054-1065.

Guerra, D., Gawlick, U., Bizarro P., 2009. An integrated

data management approach to manage health care

data. DEBS 2009

Liu Z., Krishnaprasad, M., Arora, V., 2005. Native

Xquery processing in Oracle XMLDB. SIGMOD

Conference 2005: 828-833.

Milenova B., Yarmus, J., and Campos, M., 2005. SVM in

Oracle Database 10g: Removing the Barriers to

Widespread Adoption of Support Vector Machines.

VLDB 2005: 1152-1163.

Motro, A., Yuan, Q., 1990. Querying Database

Knowledge, ACM 069791 365 5/90/0005/0173.

Oracle, 2010a. Using Continuous Query Notification,

http://download.oracle.com/docs/cd/E11882_01/appde

v.112/e17125/adfns_cqn.htm#CHDIEHHJ.

Oracle, 2010b. Semantic Technologies, http://download.

oracle.com/docs/cd/E11882_01/appdev.112/e11828/to

c.htm.

Ramamohanaro, K. and Harland J., 1994. An Introduction

to Deductive Database Languages and Systems,

VLDB Journal, 3, 107-122.

Stonebraker, M., Moore D., Brown P., 1998. Object

Relational DBMSs: Tracking the Next Great Wave,

2nd edition. Publisher Morgan Kaufmann Publishers

Inc. San Francisco, CA, USA.

Tsur, S., 1991, Deductive Databases in Action, ACM 0-

89791-430-9/91/0005/0142.

Universitaet Bonn, 2010. Active and Deductive Database,

http://idb.informatik.uni-bonn.de/research.

Venkata, K. Kanth, R., Hanckel R., Yalamanchi, A., 2008.

Using Oracle Extensibility Framework for Supporting

Temporal and Spatio-Temporal Applications. TIME

2008: 15-18.

Yalamanchi A., Srinivasan J., Gawlick, D., 2003.

Managing Expressions as Data in Relational Database

Systems. CIDR 2003.

HEALTHINF 2011 - International Conference on Health Informatics

432