Christina Engström
and Kristina Groth
CLINTEC, Karolinska Institute, 100 44 Stockholm, Sweden
HCI, CSC, Royal Institute of Technology, 100 44 Stockholm, Sweden
Keywords: Clinical decision support system, similar patient cases, terminology, structure of documentation.
Abstract: We present a field study at a surgical clinic of what data that is relevant in order to identify clinically
relevant similarities. We have observed discussion meetings in which different medical specialities decide
how to treat patients with severe diseases in the liver. Our study also includes interviews with medical
personnel, and examination of two data sources, the electronic patient records and the local quality liver
registry used within the clinic. Our findings include a model of data that can be useful when searching for
clinically relevant similarities between patient cases, as well as requirements on the functionality of an
application that can identify clinically relevant similarities.
Within a highly specialised medical care in the upper
abdominal tract at a hospital in Sweden, several
disciplines are involved in the treatment decision of a
patient. During such decision meetings, held once a
week within the area of liver, it often happens that a
doctor refers to a case experienced earlier. If no
doctor with adequate experience is present during a
meeting, or if a doctor referring to a similar case does
not remember everything, information relevant for
the diagnosis and treatment of the patient may be
missed. This implies a need for an application where
the medical information about a patient being
discussed can be compared with information from
earlier cases, information that could be useful.
Several studies have shown the need to uniform
and structure the content of medical records.
Häyrinen et al. (2008) point out that it is a challenge
to standardise content and structure of medical
journals, and that a clearly defined terminology and
uniform information structure are essential to
facilitate communication and the ability to compare
data. They argue that clearly defined terms decrease
the risk of misunderstandings, and that work-around
terminology is a requirement of applications that can
support decision making and follow-ups.
The lack of a uniform terminology makes it
difficult to reuse and communicate information since
one single term can have different meanings and
people who use the term can mean different things
(Lenz et al. 2007), which can make it difficult for
health care personnel to interpret the documentation.
This may increase the risk of wrong treatments.
Medical data stored in a structured way increases
the possibilities for processing. Considering this, an
electronic health record system (EHR) could be used
for several purposes, giving secondary advantages
(Britt 1995). The possibility to integrate a clinical
decision support system (CDSS) with an EHR
system is proposed by Porcelli & Lobach (1999).
Even though a large amount of effort has been
focusing on CDSS, it appears that only a few are in
use today. For a CDSS to be useful it is necessary to
use an existing data source where the user does not
have to manually register data for the CDSS (cf.
Grudin 1988). Also, the CDSS should not replace
manual decision making, but amplify and support the
decision process (see, e.g., Coiera 2003).
Another important issue is to make the CDSS
available at the time of the decision. O’Sullivan et al.
(2007) and Johnston et al. (1994) show that by
presenting the right information at the time of the
decision making, doctors can be supported to base
their decisions on solid grounds. They show that such
real time CDSS can enhance medical care and
Engström C. and Groth K. (2009).
In Proceedings of the International Conference on Health Informatics, pages 187-192
DOI: 10.5220/0001536901870192
support efficient exchange of knowledge between
different groups of medical staff.
In our work, we have focused on what data that
would be relevant in an application that can compare
different patient cases and present information
showing similarities. We have conducted a field
study of what data that are relevant in order to
identify clinically relevant similarities between
patients with severe liver diseases. Our study also
investigates requirements on the data sources
necessary in order to accomplish such comparisons.
Medical experts base their decisions on, e.g., existing
guidelines, case studies and experiences from
previous patients that are clinically similar to the
current case. By developing applications that use
individual patient cases as an information source
including experiences from treating these patients,
decision making in new cases can be supported
(Rossille et al. 2005, Frize et al. 2005).
A comparison of patient cases can be based on
specific data from, e.g., the electronic patient records,
or on context, e.g., by indexing documented data or
by rating different factors, e.g., earlier diagnosis
(O’Sullivan et al. 2007). Information stored as free-
text are difficult for an application to understand and
interpret, but there have been attempts to use free-
text index based on certain nouns and their definition
in terminologies like Snomed CT (Huang et al.
2003). This requires a uniform terminology since the
meaning of different terms and relations between the
terms must be clearly defined. On the other hand, it
can be a challenge to structure complex clinical
observations and store them in forms (Hogan &
Wagner 1996, Bleeker et al. 2006).
In a recent study of identifying similarities
between patient cases, Melton et al. (2006) used five
metrics to assess the degree of closeness between
cases and to discover analogous cases. They defined
similarities between patient cases by counting the
differences in characteristics. They conclude that
their models have the potential to be useful in the
area of data mining, but that they are not yet as good
as clinical experts.
In a similar study within the domain of breast
cancer, Rossille et al. (2005) propose a future CDSS
based on data warehouse and automatically getting
similar patient cases from the EHR system. They
point out that patient data today are not stored in a
format proper for automatic analysis, and that the
system architecture, therefore, must be decided first.
In a case based reasoning approach, Frize et al.
(1996) compared and identified patient cases that are
“as similar as possible”. Documented parameters
were rated, in collaboration with doctors, using
special software. They conclude that this way of
using experiences from earlier treated patients can
enhance patient handling in intensive care units.
These studies imply that an application that can
identify similarities between patient cases has the
potential to be helpful in medical care. Also, these
earlier studies point out the importance of using
uniform and structured data and terminology.
In order to understand the conditions for such an
application, we have conducted a field study at a
surgical clinic at a hospital in Sweden (Gastro). We
have identified what data that are relevant to
compare, and examined the possibilities to use
documented information as data sources.
Our research is based on close interdisciplinary
collaboration between researchers within the fields of
human-computer interaction, medical informatics,
and surgery. We have used qualitative methods and
based our work on grounded theory. The data
collection has been made through interviews,
observations, samples of medical records, and
examination of a local quality liver registry. In total
seven interviews have been conducted with liver
surgeons, a radiologist, a terminologist and a
medically trained doctor working with IT solutions.
Observations have been conducted of eight multi-
disciplinary video-mediated liver meetings, in which
decisions about how to treat the patients are made.
All interviews and observations have been recorded
and transcribed or documented with field notes (two
observations). Twelve samples of anonymised
electronic patient records from the liver decision
meeting have been analysed.
Multi-disciplinary video-mediated meetings are
held at Gastro every week within the area of liver.
The discussion during these meetings may identify
unclear issues that need to be further investigated,
focus on results from samples made and so forth, all
in order to come to a consensus about the best
possible treatment. Each patient case discussion
follows the same structure with a clinical
presentation of the patient, a walkthrough of the
radiological examinations, and a discussion of how
the patient can and should be treated.
Gastro uses an EHR system (hereafter named
TC), that supports the whole care process, and to
document different care related activities there are
different kinds of notes. These notes can, to a certain
degree, be structured by using common headlines.
The headlines have to be defined in a catalogue of
terms associated with the system, to make it possible
to search for the terms in free-text. By structuring
HEALTHINF 2009 - International Conference on Health Informatics
notes based on already defined terms the uniformity
between patient records increases.
The decision made at each liver meeting is
documented using specific notes in TC, summarising
the patient’s condition, results and interpretations of
radiology pictures, the decision made and the plan of
how the decision should be fulfilled.
Gastro also uses a local liver registry, which
includes information about all patients with tumours
in the liver, gall bladder and biliary passages, and
who has been treated with surgery. Data in the liver
registry is registered using a form based input and
consists of parts of the information that is stored in
TC, but in a more structured way. There is also
information in the liver registry that has not been
documented in TC.
4.1 Description of Patient Cases
Our observations of the liver meetings show that the
same categories of data, to a large extent, are in focus
during the discussions:
patient status, e.g., age, motivation, weight,
general health, and strength,
subjective symptoms, e.g., gastrointestinal
symptoms and tumour symptoms,
present disease, e.g., diagnosis, judgements,
treatments, progress,
present and earlier diseases and treatments,
status of the liver, e.g., the function,
examinations made and their results, e.g.,
radiological examinations, laboratory tests, and
function tests,
implemented and suggested actions, e.g.,
treatments and planned examinations.
When examining the documented notes in TC
from the decision meetings, we found that they
mainly included the categories subjective symptoms,
examinations made and their results, and
implemented and suggested actions. The most
significant difference found was that the information
in the medical records is not as detailed as in the
discussion during the liver meetings.
Examination of the liver registry indicates a
similar content, but structured using some other
patient status, e.g., sex, length, weight, BMI,
subjective symptoms, e.g., tumour symptoms,
status of the liver, e.g., the function,
examinations made and their results,
tumour information, e.g., type and size,
The category “treatments”, contains data that can be
found in the category “implemented and suggested
actions” from the liver discussions. Also, “tumour
information” is a more specific category, but is
included in the category “present disease” from the
liver discussions.
Interviews with clinically working doctors gave
an understanding of what data about a patient they
find relevant to compare to identify clinically
relevant similarities. Two of the surgeons said:
Unni: It is how old the patient is, the condition,
earlier illness, liver function, if the patient have
had earlier liver diseases, present values of the
liver function /…/ other kinds of diseases,
diagnosis and how long the disease has
Bill: It can be plenty, anatomic variations,
tumour growth in different ways, how often we
manage to do a resection, how often we can
accomplish an R0, a radical situation. It can be a
case where we have ten liver metastases that are
located a couple of centimetres deep, and we
want to make local resections. Then you get a
feeling that this is not good, we will not
accomplish an R0-situation because we seldom
do. Today, we do not have that kind of structured
documentation that can confirm these suspicions.
What is interesting is that the two surgeons stress
the level of detail that often is missing from the
medical records, e.g., how the tumour is growing,
how long the disease has proceeded, and how the
tumours are located. They also mention other
diseases as relevant, i.e, comorbidity.
During the interviews we also asked what data in
general that are needed at the liver meetings to make
a good decision about the treatment. From the
answers it is obvious that there are specific kinds of
data that are relevant for the decision, but that the
surgeons do not find relevant when searching for
similarities different patient cases:
Bill: It is data about how ill the patient is, how
he or she can manage the treatment, how
motivated the patient is to different kinds of
treatment. Then it is all the details about the
oncology treatment, how the patient has
responded to that treatment. That [kind of
information] is sometimes poorly documented.
Hence, from the interviews the following categories
have been identified as relevant content:
patient status, e.g. age, motivation, general
health, and strength,
present disease, e.g. diagnosis and how long
the disease has proceeded,
status of the liver, e.g., the function,
examination results, e.g. describing attributes
of the present disease such as anatomic
variation and localization,
implemented treatments and results,
comorbidity and earlier diseases.
The examination results appear to be an important
source of information when identifying clinically
relevant similarities. Therefore, we have focused on
relevant examinations including treatments and
diseases (e.g., comorbidity and primary disease), e.g.,
radiological examinations, laboratory test, biopsies,
endoscopic examinations, functional tests of organs,
and clinical judgements. These examinations appear
to generate the same kind of information expressed in
different ways. One such example is the function of
the liver, where there are laboratory tests that can
measure levels of certain elements in the blood, e.g.,
ASAT, ALAT, ALP, albumin, PK, bilirubin. Also,
radiological examinations can, to some extent, show
the liver function, e.g., signs of cirrhosis. There are
also functional tests that can be used to describe the
function of the liver, e.g., ICG-clearance values and
elastographic imaging.
4.2 User Needs and Requirements
Another focus in our work has been the doctors’
needs and requirements of an application that can
identify clinically relevant similarities between
patient cases. During the interviews, the clinically
working doctors had, mainly, a positive attitude
towards such an application. To exemplify what is
expected, let us take an example from one of the
interviews, in which Joe, a senior surgeon and
manager, said:
All these [radiological examinations and data
about the patient’s medical history], the
discussion and the conclusions made, that they in
the specific moment, at the same time as they are
generated, can be collected in a database. That
the database thereafter can recognise, based on
a pattern, /…/ that five patient cases look similar
and are presented.
Responses from the doctors were that such an
application should support and facilitate experienced
based care and decision making, needs to be carefully
designed, is useful only if the similarity measures are
specific enough and clinically relevant, and can be
useful during the liver meetings when there is an
uncertain patient case. Bill even said that he thinks it
could be worthwhile an extra effort of entering
information into such a system if it could prove to be
useful (cf. Grudin 1988).
However, all surgeons were not equally positive.
Unni thought that this kind of system could mainly
be useful for doctors with less experience:
It is not obvious [how such a system can be
useful] because of the way we work, with the
contacts we have, attending congresses and so
forth. We keep ourselves updated and we have a
large volume [with patients]. I think we work
pretty much like such a system without having to
use it. Such a system feels like a cookbook for
people that are not as experienced, and that can
be interesting, but I don’t think it would be useful
for me. It takes a lot of effort to build it [to fill it
with data and keep it updated].
We have also observed several liver meetings in
which associations to similar patient cases were made
in the discussions. The surgeons interviewed
responded that similar patient cases that today come
up in a discussion as a reference case are helpful for a
decision in the present case. However, such
references are dependant on the medical doctors
present during the meeting, and they are usually quite
weak since the doctors are not always able to recall
all details. Sam, a meeting participant said:
We have had a similar case earlier, exactly the
same CT. /…/ We have the answer but I cannot
remember exactly.
Nora, a radiologist said in the interview that she
had worked with a similar application, in which
similarities were based on registration of ten codes
per patient, but that the similarities were not specific
enough. She found such an application useful, if the
similarities are more specific and detailed.
When asking Bill in what situations such an
application could be useful he said:
It is in those cases that are uncertain, for
example if we should do an operation or not. In
many cases it can be questioned if it is
meaningful to do an operation from a tumour
biological approach. If we had a fine grained
database that could show the results of these
kinds of patients.
It appears that the granularity of the data that are
compared and that are presented from similar cases is
an important aspect for success.
4.3 Documentation Routines
One important part in this kind of application is the
existing routines and the doctors’ attitude towards
documentation and changes of documentation
routines, including the terminology used.
All doctors interviewed agreed that the
terminology is important in the medical work. Some
HEALTHINF 2009 - International Conference on Health Informatics
said that they are careful how they express
themselves, to avoid misunderstandings. They are
aware of a certain degree of shifting in the
terminology used, but think that misunderstandings
are rare since the situation provides a context for the
understanding. They said that each doctor has her
own way of expressing herself, but the use of terms
should not be restricted. Jim, a senior surgeon, said
that what terms that are used and how they are used
is a negotiation based on different aspects:
It is not that you try to talk using the terms
defined in TC. Instead the terms in TC needs to
be adapted to how people express themselves.
Then, of course, when you document you still
need to adapt [to the term catalogue in TC],
everybody can’t express themselves as they want
and all terms can’t exist. The goal with the terms
is that you can use them to search.
In TC there is a term catalogue that can be used to
increase the ability to search and unify the
documentation. However, the interviewed surgeons
found this work somewhat complicated because of a
lack of administrative routines for defining new
terms, both on a clinic and hospital level. Earlier
attempts to apply for new terms to be added to the
term catalogue had no result. The surgeons are also
unaware of new centralised efforts of a terminologist,
i.e., their attempts to work with terminology issues
have been down prioritised.
After the liver meetings each decision is
documented in TC. The decision note is not a
formalised structure in TC, but it follows a template
with headings such as decision and activity, how to
carry out the activity, major diagnosis, and whether
the decision involves operation or not. The
formalised way of doing this would be to use the
term catalogue to build the structure inside TC. Bill
said that in the beginning he found the template for
the decision documentation too structured, but now
he would like it to be more structured. He would like
the decision itself to be documented in more detail,
not only suggesting one alternative, but also why
other alternatives have been rejected. Bill also said
that one reason why this often is missing could be
that the surgeon documenting the decision is usually
not so experienced and, therefore, may not have
followed the whole discussion.
The local liver registry and TC consist, to a large
extent, on the same kind of information. The main
difference is that the documentation in TC is more
specific and detailed, but the liver registry contains
some specific parameters for statistical analysis, e.g.,
the amount of bleeding during the operation, if the
circulation was turned off to the liver during the
operation, how the liver was cut, how much of the
liver that was removed, and what kind of operation
that was made. About the liver registry, Bill said:
It is that kind of data that you want to use to
learn for the future, what patients that can
manage a turndown of the vessels or not. These
kinds of things are never documented in TC. /.../
There is usually an operation documentation, but
in the best case the operation code coincides
with what was actually done.
Based on our observations, interviews and
investigations of documentation in TC and the liver
registry, we identified five categories of information
that can be relevant when identifying clinically
relevant similarities between liver patients:
General Data about the Patient, including
general health and strength, age, function of
other organs, and clinical assessment.
Data about the Liver Function, including
laboratory tests, examinations of the liver
function, and comorbidity concerning the liver.
Data about the Present Disease, including
radiological examinations describing the type of
examination, contrast load, position and size of
affected tissue, and relationships to large blood
vessels and bilary passages, laboratory tests,
biopsies, and diagnosis.
Data about the History of Diseases, including a
documented history of diseases with diagnose,
spread of disease and point of time.
Data about Treatments, including point of time,
kind of treatment, and effect.
The above presented data model is based on
information that is documented in TC and the liver
registry, but origins from mainly three sources:
examinations, documented medical history including
earlier examinations, treatments and diseases, and the
patient’s subjective description of the symptoms,
motivation, disease history and so forth.
One consideration regarding the data model
concerns the validity of different data, e.g., the
documented symptoms, which are dependent on the
patient’s communication ability. This results in a
subjectivity, which makes comparisons of symptoms
unreliable. Symptoms were not mentioned by the
doctors during the interviews, but were frequently
discussed during the observed liver meetings.
Although symptoms can give important information
about the disease and possible treatments, we have
chosen not to include such data in the proposed data
model because of this uncertainty and subjectivity.
One interesting question concerns how the data of
interest can be found and should be described.
We have shown that there are several alternatives
used to describe the same kind of information. To
make it possible to compare data, it must either be
documented in the same format and/or using the
same terms, or it must be possible to use a translator
between the different formats and/or terms.
During the interviews the radiologists expressed
that they are careful in how they express themselves,
both during the meetings and when documenting the
examination. If they say that they are certain that the
dark change on the CT is a tumour, then they are.
Otherwise, they say that it appears to be or behaves
like a tumour. It is a relevant difference between a
certain tumour and a likely tumour and an application
should be able to understand such difference. The
question is how the level of uncertainty can be used
when searching for similarities. One solution could
be to use same kind of rating variables, but needs to
be further examined.
Also, the level of detail in content is important to
make the application useful for experienced doctors.
If the data is not detailed enough, it will not identify
similarities on the right level.
We have focused on what data that would be relevant
in an application that can compare different patient
cases and present information showing similarities.
This is only part of a broader perspective, including
not only data but also information from different
kinds of media.
The aim of such an application would be to create
a kind of “clinical binocular” that can focus on the
right information at specific moments. It should not
only cover for situations when the right experience is
missing, but also for situations when a doctor may
not fully remember the previous case or when
individual interpretations previous cases influences
what is remembered. In some sense it should
strengthen the “clinical eye”.
It is also of interest to keep in mind that the
functionality of such an application affects the
content of the data sources, i.e., it needs to be
detailed, searchable, and structured. This, in turn,
affects how the documentation is made, something
that may require changes in documentation activities
and routines. Two important issues have been
pointed out during the interviews: the effort to make
changes must be rewarded and give a clear surplus
value, and the changes must be easy to implement
with helpful assistance and aid that reduce the
doctors’ efforts.
We are grateful to all medical personnel who have
been there for us when conducting our studies.
Bleeker S., Derksen-Lubsen G., van Ginneken A., van der
Lei J., & Moll, H. (2006). Structured data entry for
narrative data in a broad speciality: Patient history and
physical examination in pediatrics. BMC Medical
Informatics Decision Making, 6 (29).
Britt H., Beaton N., & Miller G. (1995). General practice
medical records: Why code? Why classify? Australian
Family Physician, 24 (4), 612–615.
Coiera E. (2003). Guide to health informatics. second
edition, Arnold.
Frize M., Solven F.G., Stevenson, M., & Nickerson, B.G.
(1996). Computer-assisted decision support systems for
various medical applications. Proc. of 10th Int. Conf. of
Women Engineers and Scientists, 91–93.
Grudin J. (1988). Why CSCW applications fail.
Proceedings of CSCW’88, ACM.
Hogan W. R., & Wagner M. M. (1996). Free-text fields
change the meaning of coded data. Proceedings of
AMIA Symposium, 517–21.
Huang Y., Lowe H. J., & Hersh W. R. (2003). A pilot study
of contextual UMLS indexing to improve the precision
of concept-based representation in XML-structured
clinical radiology reports. JAMIA, 10 (6), 580–587.
Häyrinen K., Saranto K., & Nykänen P. (2008). Definition,
structure, content, use and impacts of electronic health
records: A review of research literature. Int J Medical
Informatics, 77 (5), 291–304.
Johnston M. E., Langton K. B., Haynes R. B., & Mathieu A.
(1994). Effects of computer-based clinical decision
support systems on clinician performance and patient
outcome: A critical appraisal of research. Ann. Intern
Med., 120, (2), 135–42.
Lenz R., Beyer M., & Kuhn K. A. (2007). Semantic
integration in healthcare networks. Int J Medical
Informatics, 76 (2- 3), 201–207.
Melton G. B., Parsons S., Morrison F. P., Rothschild A. S.,
Markatou M., & Hripcsak G. (2006). Inter-patient
distance metrics using SNOMED CT defining
relationships. J Biomedical Informatics, 39 (6), 697–705.
O'Sullivan D., McLoughlin E., Bertolotto M., & Wilson D.
C. (2007). Mobile case-based decision support for
intelligent patient knowledge management. Health
Informatics J., 13 (3), 179–193
Porcelli P. J., & Lobach D. F. (1999). Integration of clinical
decision support with on-line encounter documentation
for well child care at the point of care. Proceedings of
AMIA Symposium, 599–603.
Rossille D., Laurent J. F., & Burgun A. (2005). Modelling a
decision-support system for oncology using rule-based
and case-based reasoning methodologies. Int J Medical
Informatics, 74 (2-4), 299–306.
HEALTHINF 2009 - International Conference on Health Informatics