Knowledge-based System for Urinalysis

Fabrício Henrique Rodrigues, José Antônio Tesser Poloni, Cecília Dias Flores and Liane Nanci Rotta

Universidade Federal de Ciências da Saúde de Porto Alegre, Porto Alegre, Brazil

Keywords: Ontology, Knowledge-based System, Knowledge System, Expert System, Bayesian Networks, Health

Informatics, Urinalysis.

Abstract: Urinalysis is a very important test of laboratory medicine, providing valuable information about

metabolism, kidney, and urinary tract. For several reasons, including lacking of professional qualification, it

does not receive the proper attention, what prevents it to achieve its whole power. Considering that, a

knowledge-based system for decision support in urinalysis could help to change this situation, being useful

to professional training, decision support during the process or even the automation of the test. This paper

proposes the development of such a system, employing ontologies, Bayesian networks, and templates of

cognitive tasks to treat domain knowledge. Then, urinalysis is briefly discussed and system architecture is

presented, as well as the current state of the work and future steps.

1 INTRODUCTION

Urinalysis is probably the earliest test of laboratory

medicine (King, Strasinger, 2008). It can be defined

as the testing of urine with procedures commonly

performed in an expeditious, reliable, accurate, safe,

and cost-effective manner (CLSI, 2001). Nowadays,

it is an integral part of the patient examination and is

composed by the following main steps:

 Urine collection and storage: obtains the sample

and stores it until the analysis itself;

 Direct Observation: examines colour, turbidity,

and odour of urine. It is falling out of favour in

light of technological advance;

 Physicochemical analysis: carried out by means of

dipstick – a plastic strip with reactive areas that

gives an approximate estimation of some physical

and chemical parameters of urine (e.g. density, pH,

albumin). This estimation is detected through

colour change of the respective reactive areas;

 Microscopy: it is done over a spot of urine in a

microscope slide and identifies the particles in it

(e.g. cells, crystals, microorganisms), sometimes

using some auxiliary tools (e.g. polarized light,

sediment stains). The same slide is analysed tens

of times, in different microscopic fields (i.e.

regions of the slide). After each field analysis, the

observed findings are registered.

Even though inexpensive and dealing with an easily

collected body fluid, urinalysis is a very important

test. It can provide valuable information about many

of the body’s major metabolic functions, as well as

the condition of the kidney and urinary tract (King,

Strasinger, 2008).

However, in spite of its importance, this

laboratory exam has not received the proper

attention, what prevents it to achieve its whole

power. One of the main expressions of this is that,

generally, the urinalysis is too focused on the

physicochemical analysis, leaving microscopy to a

secondary role, being performed without correct

methods, equipment, and professional qualification.

This way, the reported results relies too much on an

approximated examination of physicochemical

parameters, with significant particles being missed

or misinterpreted in microscopy – which means

missing valuable information about the patient

(Fogazzi, Verdesca, Garigali, 2008).

In order to change this scenario, (Fogazzi,

Verdesca, Garigali, 2008) point out the following

requirements:

i. Use of correct method for patient preparation

and urine collection and handling;

ii. Capability to identify the most important

particles in urine;

iii. Knowledge of clinical meaning of the urine

particles;

iv. Capability to arrange urinary findings in a

clinical context.

Except for (i), all the given requirements are about

514

Rodrigues F., Poloni J., Flores C. and Rotta L..

Knowledge-based System for Urinalysis.

DOI: 10.5220/0004952305140519

In Proceedings of the 16th International Conference on Enterprise Information Systems (ICEIS-2014), pages 514-519

ISBN: 978-989-758-027-7

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

pure cognitive and informational tasks, which may

be suitable to computational modelling. Such

requirements reveal three different facets that are

necessary to take into consideration in order to

rightly portrait the domain – with specific

representational tools suitable for each of them.

The first facet is the representation of the

complexity of the concepts involved in the task. This

aspect is present in the information needed both to

recognize particles in urine and define their clinical

meaning as well as to describe all the findings and

their clinical contexts. For that it may be useful to

employ ontologies. An ontology can be defined as

an explicit specification of a conceptualization

(Gruber, 1993). Generally, it is represented as a set

of concepts, a set of relations among these concepts,

a set of attributes to describe them, and other axioms

about the conceptualization.

The second facet is the uncertainty inherent to

medical domain (Schwartz and Elstein, 2008).

Regarding urinalysis, it is mainly due to the non-

deterministic nature of the relations between

findings and clinical contexts (i.e. single particles or

sets of findings are not always related to the same

clinical condition and may be arranged in different

clinical contexts) and the usual incompleteness of

information (i.e. not all findings that characterize a

clinical context are always present at the same time

at the same sample). Such uncertain aspect of the

domain can be dealt with using Bayesian networks

(BN). BNs are directed acyclic graphs in which the

nodes represent domain variables and the arcs

represent influences among these variables (Pearl,

1985), quantified by conditional probabilities tables.

The last facet is the reasoning processes and

heuristics needed to relate findings and decide what

to do to next during the sample analysis. Even

though the characteristics of particles are known as

well as their clinical meanings and associations, it is

still needed further cognitive skills to take advantage

of that knowledge (e.g. selecting a tool to identify a

particle, indentify inconsistencies in the findings).

With the aim of modelling so, we can use task

templates (i.e. reusable combinations of model

elements, that supply inferences and tasks typically

used to solve problem of a particular type)

(Schreiber et al, 1999). Examples of these tasks are

diagnosis, prediction and monitoring.

Thus, considering:

 The importance of urinalysis;

 The requirements to be met in order to allow

urinalysis to reach its whole power;

 The cognitive and informational nature of such

requirements and;

 The existence of computational techniques and

artefacts suitable to modelling them;

it seems to be possible and useful to develop a

computational system with a representation of the

domain of urinalysis able to fulfil the requirements

earlier mentioned. Encompassing such capabilities,

this system could be adapted and used for a variety

of purposes – such as professional training, decision

support and urinalysis automation.

Following this hypothesis, this paper proposes

the development of a knowledge-based system – i.e.

that uses artificial intelligence techniques in

problem-solving processes to support human

decision-making, learning, and action (Akerkar,

Sajja, 2009) – for the domain of urinalysis. Due to

the exposed, the core of the system is planned to be

built using ontologies, BNs and template tasks, each

being used for modelling the respective facet of the

domain. In spite of the different possible uses for it,

as a first version, the system is being conceived as a

decision support tool that will be used to:

 Answer questions about the domain;

 Evaluate user’s hypotheses;

 Guide the user during the exam (i.e. user provides

new findings to system evaluation, which returns

expert advice to the user).

The rest of the paper is organized as follows: section

2 presents urinalysis knowledge needs, section 3

presents the proposed system architecture, section 4

presents the current state of the work and next steps,

and section 5 brings conclusions.

2 KNOWLEDGE NEEDS

In order to develop a system aiming to help in

urinalysis, it is imperative to review its context and

knowledge needs. As discussed before, executing a

good urinalysis is in great extent a matter of doing a

good microscopy. In this way, knowledge about the

fine distinctions between types of particles in urine

is unquestionably mandatory for a good urinalysis.

Yet, since urinalysis is an indirect way of

assessment of patient condition, achieving this

objective also depends on the knowledge of the

clinical conditions that can affect the urine

composition – what is important not to miss relevant

particles and to correctly interpret them.

These conditions may represent rather intricate

processes, albeit their influence over urine is much

more limited. Then, with the purpose of avoid

unnecessary complexity and still improve analysis, it

is possible to use urinary profiles – i.e. combinations

Knowledge-basedSystemforUrinalysis

515

of urinary findings associated to clinical conditions

(Fogazzi et al., 2008) – to summarize such influence.

There are profiles for a variety of conditions,

including nephritic/nephrotic syndromes, urinary

tract infection, and hepatic disorders.

Yet, in order to perform a good urinalysis, it is

not sufficient to know only the clinical and visual

aspects of particles. It is also necessary to better

know the whole urinalysis process. This is due to

the fact that there are a number of conditions and

events (some of them prior to urinalysis itself) that

can influence urine contents, leading it to

misrepresent the real patient condition or hindering

such information. These conditions and events were

extracted from literature and interviews with an

expert. Some of them are listed below:

 Sample contamination (e.g. patient’s lack of

hygiene during collection, remains of cleaning

substances in collection bottle, women’s period,

intentional urine diluting );

 Exposure to heat or light during storage, that may

degrade some substances and particles;

 Influence of conservation methods (i.e. chemical

preservatives and refrigeration)

 Sample of urine that is too much pigmented (e.g.

due to some medicine patient is having) painting

dipstick, masking the real color change due to

chemical reaction

 Confusing particles (i.e. different particles with

similar morphology and visual aspects)

Besides urinary profiles and misleading urinalysis

events, it is still important to urinalysis professional

some additional punctual knowledge about clinical

conditions, mainly related to their expression in

other laboratory exams (e.g. blood). This allows

verifying the result of such exams that the patient

may also have been subject of (when such

possibility is available) or even directly enquiring

patient’s physician. Such information may be useful

to evaluate some hypothesis about a clinical

condition that some unusual urinary finding has

pointed out – and thus be able to verify if the finding

is genuine or caused by an error.

Finally, there are a lot of actions that may be

taken during urinalysis (e.g. use of a specific kind of

microscopy, add some substance to microscope

slide, ask new sample collection, inspect patient’s

urinalysis history available in laboratory). Mastering

the context in which each action should be taken is

another requirement for a good urinalysis.

Considering the urinalysis context presented

here, it was identified the following main reasoning

tasks performed during urinalysis:

 Assess quality of a sample and decide whether or

not ask for new collecting;

 Classify sample in an urinary profile;

 Use acquired information to guide search and

interpret new findings;

 Formulate hypotheses and use additional

information to confirm or refute them;

 Identify incoherencies among findings;

 Identify problems/errors during the test;

 Decide which action to carry out next;

Briefly, these are the knowledge requirements for a

good urinalysis. Consequently, it should be observed

in order to develop a system with a meaningful

knowledge model and that provides effective

guidance to the user during urinalysis. The next

section presents the system architecture proposed to

fulfill those requirements.

3 PROPOSED ARCHITECTURE

As it was presented, urinalysis domain involves a

great number of concepts (e.g. particles, substances,

profiles, tools), with a strong descriptive aspect (e.g.

visual aspects) and many kinds of associations

between them. Taking it into account, we decided to

have an ontology as the main knowledge source for

the system. It is intended to cover the domain

approaching three main aspects:

 Urine representation: include all particles and

substances that can appear in urine, with their

relations and visual attributes, as well as all of

physicochemical parameters of urine;

 Clinical information: include the urinary profiles,

additional information about selected clinical

conditions and other laboratory tests, and a model

of patient, with its key characteristics (e.g. gender,

age);

 Urinalysis process: include the representation of

the dipstick and all its physicochemical tests and

their relation with the respective substance or

parameter, all the tools and actions the analyst can

take, the events and situations that can change

urine composition, and associations between

findings

Bearing in mind the existence and relevance of

uncertainty in health domain, the system will also be

composed of BNs. But, in view of its complexity in

building and maintenance (Koller, Pfeffer, 1997),

instead of a large BN, the system will have specific

small BNs for those portions of the domain that are

more sensible to uncertainty. As much as possible,

the nodes of BNs will have correspondence to

concepts of the ontology, so as to guarantee further

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

516

interpretation of the conclusions taken from BNs in

terms of ontology axioms.

Over these knowledge models, it will be

developed a system with three layers, as shown in

figure 1. The first is the interaction layer. It is

designed to enclose the modes of interaction

between user and system (e.g. patterns for questions,

evidence and hypothesis providing, system answer

and guidance). The interaction will be all based on

ontology concepts. Auxiliary, it will also be used a

lexicon for the concepts that can be referred to by

different terms and a disambiguation mechanism for

terms that can be mapped to more than one concept.

Figure 1: System architecture.

The next is the oracle layer. This layer is designed to

receive questions from interaction layer in one of the

predefined patterns and formulate the appropriate

queries (either for ontology or some BN) to get the

answer, returning it to interaction layer.

Analogously, it is also designed to receive

hypotheses about a sample (e.g. combinations of

findings and a urinary profile user thinks that is

compatible to each other), also in a predefined

pattern, evaluate whether it is true or false (using

ontology) or the likelihood of its truth (using some

BN) and return the result to interaction layer.

Finally, there will be the expert layer. This layer

is designed to simulate the expert behaviour during

the exam. Thus, it is intended to be used to evaluate

any new information provided by the user (which

will be received through some pattern of interaction

layer) and to perform the reasoning tasks enumerate

in the previous section, always considering all the

information already gathered during sample

analysis. Aiming to ease the development of this

layer, as well as to make it more understandable and

maintainable, it was conceived as a chain of five

task templates extracted from (Schreiber et al.,

1999): monitoring, diagnosis, classification,

prediction and assessment. The interactions among

them are presented in figure 2.

Figure 2: Task Template Chain.

Monitoring is the task of analyzing an ongoing

process to verify if it behaves according to the

expectations. It gets as input historical data about a

monitored system and gives as output any found

discrepancies from the expected values, with no

further investigation of its causes. The task starts

receiving new findings and evaluating its parameters

against some norms. The difference is, then,

classified as a type of discrepancy or as normal case.

In our system, monitoring will be used to look

for inconsistencies among the findings (e.g. acid

crystals found in alkaline urine), signs of false-

positives and/or false-negatives (e.g. high levels of

blood found on dipstick but no blood cells in

microscopy) or other problems in the sample (e.g.

too much epithelial cells). It will be based on

ontological knowledge as well as some trigger rules

learned from expert. All already gathered data will

be considered. Moreover, a time index will be used

to judge possible discrepancies (i.e. some

discrepancies will be only considered so if

discrepant values persist after the analysis of a given

number of microscopic fields). As output, it will be

returned any found inconsistency or sign of false-

positive/false-negative or other sample problem.

Diagnosis means finding the fault that causes

some malfunction in a system. The inputs of this

process are symptoms and the outputs are the fault

and evidences found of it. It generally uses a model

of the behavior of the system being diagnosed.

Diagnosis starts by taking the complaints and

making hypotheses about the problem by going

backwards in a causal network. Then, the actual

findings are compared with the signs that should be

observed for each hypothesis, excluding those in

conflict with the findings. The remaining hypothesis

and the observations that led to it are the output.

In the proposed system, diagnosis will be done

over the output of monitoring, inferring the roots of

any identified false-positive/false-negative result,

problem or inconsistency. Hypothesis generation

will be based on ontological knowledge and there

Knowledge-basedSystemforUrinalysis

517

will be a BN to evaluate the likelihood of concurrent

hypotheses, when more than one remains at the end.

The causes of the discrepancies (or the possible

ones, ordered by likelihood) will be given as output.

Classification task represents the establishment

of the correct category of an object available for

inspection, based on its characteristics. As input, it

takes an unclassified object and gives one or more

classes as output. It is done by taking candidate

classes and matching their attributes with those of

the object. As some attributes of the object conflicts

with one of the candidate class, this is discarded.

According to the matches, none, one or more than

one classes can remain as the output.

The system proposed will use classification to

identify the urinary profile(s) of the sample in

accordance with all the data already gathered –

including the output of diagnosis task. Profile

definition will be ontology-based. In addition, given

that it is not usual to find all the findings needed to

unambiguously point to a single profile, a BN will

be devised to indicate the likelihood of the

alternative hypotheses. Analogously, classification

task will be used to classify particles whose visual

attributes are identified, but the type is not

recognized by the user. In the same way, particles

will be described with ontology concepts and a BN

will be used to evaluate alternative hypotheses.

Prediction is the task of analyzing current system

behavior to infer a description of system state in

some point of the future. For that purpose, it uses a

model of system behavior.

This task is planned to be used to tell what

findings are likely to be seen when analyzing the

next microscopic field of the microscope slide. It

will use all data already gathered and the outputs of

diagnosis and classification tasks. The main tool for

this task will be a BN calibrated to indicate the

probability of the presence of each particle in the

sample. Even though it is not exactly the canonical

use of the prediction, since predicting particles to be

found does not represent a future state of the sample,

we believe that the analogy is valid and that the

general idea will be useful to our case as well.

The goal of assessment task is to find a decision

category for a case, based on domain specific norms

(i.e. heuristic rules). The input is data about the case

and, sometimes, case-specific rules. The output is a

decision category. It starts receiving a case and

selecting a norm to evaluate it. The evaluation

involves both case features and the available classes

of decision. Depending on the result of norm

evaluation, a decision class may be chosen. If it is

not possible to select a decision class, another norm

is evaluated. Sometimes more than one norm match

is needed to assess a decision class.

We will use assessment to define what to do next

in the analysis, chosen from a possible list of

actions, including (but not limited to) the use of

some tool (e.g. a special kind of microscope),

searching for additional information (e.g. patient

history) and/or a specific particle in the slide. The

case representation will include all the information

already gathered and the output of all other tasks.

This task will be largely based in heuristics to be

learned from expert. The possible actions will be

described in terms of ontology concepts. As

sometimes a sequence of actions may be needed,

some planning routine may be run during this task.

4 STATE OF THE WORK

Several interviews with a urinalysis expert and a

literature review about the domain are already made.

With this material it was possible to devise the

project of the system (which was briefly presented in

this paper) and a plan of execution. In addition, it

was already registered about 90 competency

questions to guide ontology development, over 250

ontological concepts, up to 30 types of ontological

relations and attributes, plus dozens of heuristics

used by expert during the analysis.

Presently, we are working to formalize the

concepts and to structure the ontology using the

Unified Foundational Ontology (UFO) (Guizzardi,

Wagner, 2005), which was chosen due to its strong

logical framework and its cases of success. The

ontology is going to be implemented using the Web

Ontology Language version 2 (OWL2) (Grau et al,

2008). OWL2 was chosen in view of its status of

W3C recommendation, which favors its stability,

and the set of tools built based on it, including a

powerful ontology editor – the Protégé OWL (Rubin

et al, 2005). The limitations in expressiveness of

OWL2 in comparison to UFO are already being

considered in order to be mitigated.

Following, we are going to develop the BNs,

with its structure based on the ontology model and

the probabilities calibrated by the expert. The

resulting BNs will be implemented using UnBBayes

framework (Matsumoto et al, 2011). Next, we are

going to adapt the mentioned cognitive task

templates to urinalysis domain. After that, we will

work on the interface layer. All software artifacts

will be developed in Java.

Finally, we plan to validate our work in two

ways: (i) confronting ontology and BNs

ICEIS2014-16thInternationalConferenceonEnterpriseInformationSystems

518

individually, as well as the whole system, against

real urinalysis cases, in order to verify if they are

able to reach correct conclusions in comparison with

expert performance and (ii) providing the system to

urinalysis professionals and students and evaluate its

effectiveness as training and decision support tool

(i.e. whether or not it improves their capabilities).

5 CONCLUDING REMARKS

Urinalysis is a relatively inexpensive but powerful

and very important laboratory test, commonly

employed in patient examination. In spite of that, for

numerous reasons, it has not received the needed

attention to achieve its whole power, which has its

roots mainly on the lack of professional qualification

and insufficient knowledge about the test.

Given that most of the problem relies on

cognitive tasks, suitable to computational modelling,

it was formulated the hypothesis that it is possible to

develop a knowledge-based system representing

urinalysis domain and that this system can be useful

to enhance this laboratory exam. We also believe

that this usefulness can be materialized in many

ways – contributing to professional qualification,

decision support and even to urinalysis automation.

With the aim of testing such hypothesis, this

paper proposes the development of such a system.

To achieve this objective, we decided to use

ontologies to model the domain, BNs to treat its

uncertain aspect and task templates to formalize the

reasoning tasks needed to well perform the analysis.

After literature immersion and interviews with an

expert in the domain, the system was designed and

the ontology construction has started.

Even though being an apparently simple

analysis, dealing with a single and so trivial body

fluid, urinalysis revealed itself as a rather complex

area. This challenging nature can be exemplified by

the great amount of concepts involved (over 250,

selected from an initial list of about a thousand of

them), by the intense flow of information during the

analysis, and the resultant intricate heuristics needed

to treat it – which demanded a handful of task

templates to represent. It is indeed a domain that

would certainly take several years to be mastered by

a novice professional.

Nevertheless, precisely due to that challenging

nature of the domain, the importance of this work

grows stronger. Besides an intelligent system to

support urinalysis, accepting different interface

layers according to the intended use, it is also being

developed an ontology model that has valuable in

itself. This model may serve as base for a series of

useful applications for the domain, not even

imagined yet. Still, considering that the

methodological knowledge to be developed during

this work may extrapolate its domain, our work may

serve as guidance to similar initiatives in correlate

domains, such as other laboratory exams.

REFERENCES

Akerkar, R., Sajja, P. 2009. Knowledge-Based Systems,

Jones and Bartlett Publishers. Sudbury, US, 1

edition.

Clinical and Laboratory Standards Institute (CLSI), 2001.

Approved Guideline GP16-A2: Urinalysis and

Collection, Transportation, and Preservation of Urine

Specimens, CLSI. Wayne, Pa., 2

edition.

Grau, B. C., Horrocks, I., Motik, B., Parsia, B., Patel-

Schneider, P., Sattler, U., 2008. OWL2: The next step

for OWL. In: Web Semantics: Science, Services and

Agents on the World Wide Web 6.

Gruber, T. R., 1993. A Translation Approach to Portable

Ontology Specifications. In: Knowledge Acquisition.

Guizzardi, G.; Wagner, G., 2005. Some applications of a

unified foundational ontology in business modelling.

In: Business systems analysis with ontologies, Idea

Group.

Koller, C., D., Pfeffer, A., 1997. Object-Oriented Bayesian

Networks. In: 13th Annual Conference on Uncertainty

in AI (UAI).

Matsumoto, S., Carvalho, R.N., Ladeira, M., Costa,

P.C.G, Santos, L.L., Silva, D., Onishi, M., Machado,

E., 2011. UnBBayes: a Java Framework for

Probabilistic Models in AI. In: Java in Academia and

Research. iConcept Press Ltd.

Pearl, J., 1985. Bayesian Networks: A Model of Self-

Activated Memory for Evidential Reasoning. 7

Annual Conference of the Cognitive Science Society.

Schreiber, A. T., Akkermans, H., Anjewierden, A., de

Hoog, R., Shadbolt, N., Van de Velde, W., Wielinga,

B., 1999. Knowledge engineering and management:

the CommonKADS methodology, MIT Press.

Cambridge, MA.

Schwartz, A., Elstein, A. S., 2008. Clinical reasoning in

medicine. In: Clinical Reasoning in the Health

Professions. Elsevier, 3

edition.

Rubin, D. L., Knublauch, H., Fergerson, R. W., Dameron,

O., Musen, M. A., 2005. Protégé-OWL: Creating

Ontology-Driven Reasoning Applications with the

Web Ontology Language. In: AMIA Annual

Symposium Proceedings 2005. American Medical

Informatics Association.

Knowledge-basedSystemforUrinalysis

519