relations

[GoogleCodeExporter]

I was trying to find out how to relate for example signs with their
disease. My use case is being able to say for example "generalized motor
manifestation is a sign of seizure". 
There will of course be several signs, for example loss of consciousness,
or history of unconsciousness.

I didn't find any relation in the current OGMS file, and I was wondering
how to link the different elements?

Thanks,
Melanie

Original issue reported on code.google.com by mcour...@gmail.com on 11 Jan 2010 at 11:54

[GoogleCodeExporter]

I think we need sign_of and symptom_of as relations.  I think the biggest 
question is
to what to make the domain and ranges of these relations.

X sign_of Y (is X a sign report, is Y a disease?)
X symptom_of Y (is X a symptom report, is Y a disease?)

Any thoughts?

Original comment by albertgo...@gmail.com on 12 Jan 2010 at 5:46

[GoogleCodeExporter]

don't see why we need to replicate the classes in the relation name. Won't one 
relation do?
What can we infer from such a relation?

Original comment by alanruttenberg@gmail.com on 12 Jan 2010 at 7:43

[GoogleCodeExporter]

I would suggest linking bodily features to potentially disease.

Arguments for bodily features: generalized motor manifestation could be a
manifestation of a disease, a sign or a symptom. I don't think it could be a
phenotype though, but there may be case where we need to link the phenotype to 
the
disease?
In my use case, I want to be able to query both ways, i.e. "what type of 
diseases
have motor manifestations" (whether those are signs/symptoms/manifestations) 
and also
what are the bodily features associated with seizures, so I think that would 
support
having the domain to bodily features.

I am uncertain about the range - I think we want to link 
signs/symptoms/manifestation
to the disease, and that the disorder is just the physical basis of the 
disease. So
for example I want to say "motor manifestation" when observed by physician is a 
sign
of seizure, seizure being the disease, and abnormality in the brain being the
corresponding disorder?

Original comment by mcour...@gmail.com on 12 Jan 2010 at 8:00

[GoogleCodeExporter]

Except for pathognomonic signs like a Cardinal's cap in neurological anthrax, 
any
relation between signs/symptoms and disease would have to be at the instance 
level if
we are to adhere to the principle of universality.

Not every instance of influenza disease is associated with an instance of 
fever, and
certainly not vice versa.

Not all appendicitis has abdominal pain, nausea, vomiting, elevated WBC count, 
etc.

The diagnostic challenge arises in part because symptoms and signs are so
non-specific for given diseases.

Original comment by hoga...@gmail.com on 12 Jan 2010 at 8:30

[GoogleCodeExporter]

Bill...I think we abandon the principal of universality for all terms under the
'defined classes' hierarchy (and any relations between instances that use 
defined
classes for the domain or range).  But just like the FMA is a canonical 
reference for
anatomy, shouldn't there be canonical references for symptomatology (or
approximations thereof)?  If not, what are these resources trying to capture:

https://health.google.com/health/ref/
http://symptoms.webmd.com/
http://www.mayoclinic.com/health/symptom-checker/DS00671

Alan...what I would be interested in is an answer to a sparql query such as

@prefix ogms: <http://purl.obolibrary.org/obo/ogms.owl> .
select ?symptom
where ?symptom ogms:symptom_of Brucellosis

which returns a subset of those entries for Brucellosis listed on (say) Google 
Health:
    * Fever
    * Chills
    * Excessive sweating
    * Weakness
    * Weight loss
    * Fatigue
    * Headache
    * Abdominal pain
    * Back pain
    * Loss of appetite
    * Joint pain

The subset would be those that conform to some sort of OGMS principle for being 
the
content of a symptom report.  Is there a way to work this out coherently?

Original comment by albertgo...@gmail.com on 12 Jan 2010 at 8:50

[GoogleCodeExporter]

[deleted comment]

[GoogleCodeExporter]

_Bill...I think we abandon the principal of universality for all terms under the
'defined classes' hierarchy (and any relations between instances that use 
defined
classes for the domain or range)._

OK, I was not aware of that.

_But just like the FMA is a canonical reference for
anatomy, shouldn't there be canonical references for symptomatology (or
approximations thereof)?_

In other words, canonical abnormality?  Or what is canonically abnormal in
Brucellosis?

_If not, what are these resources trying to capture?_

It's hard to say unless the resource is upfront about exactly what it is trying 
to
capture.  Some resources are making a long laundry list of any symptom no 
matter how
uncommon and some resources are doing what you say, that is, trying to describe 
a
typical case of, for example, Brucellosis.  

But the resource almost never tells you their criteria for how common a symptom 
must
be, to be included in the 'typical' list.

For sure, the 'typical' or 'canonical' case of Brucellosis won't have all those
symptoms on presentation for medical care.

Original comment by hoga...@gmail.com on 12 Jan 2010 at 11:09

[GoogleCodeExporter]

I agree, most diseases will not have necessary conditions but instead are 
diagnosed based on sufficient 
conditions which in turn are based on probability of them happening together. 
Of course, there are exceptions 
to this:
   - Bill gave the example of Cardinal's cap in neurological anthrax
Other examples are:
   - Grey-Turner's sign and Cullen's sign in Acute pancreatitis
   - Janeway lesions are pathognomonic of Infective Endocarditis. However, the differential diagnosis for it can 
include other infective diseases like Typhoid or autoimmune diseases like SLE. 

So, should the model of a disease capture the common clinical features 
associated with it even though they 
may not all be present in every instance of the disease? A subset of the 
clinical features are sufficient to 
diagnose the disease - perhaps the diagnosis part can be captured as a 'rule' 
or a set of 'rules' instead. 

Going back to the original question posed by Melanie on the type of relation to 
be used, there are two 
candidates that I can think of 
       1. 'manifestation_of'  -  implying that clinical features result from the realization of a disease.
       2. 'part_of' - implying that clinical features are a part of the disease process. 

Original comment by sivaram....@gmail.com on 13 Jan 2010 at 3:13

[GoogleCodeExporter]

The distinction between ontological (universal) assertions, even for so-called
defined classes, and probabilistic assertions that make up general medical 
knowledge,
should help to avoid an impossible scope for OGMS - i.e. it is highly advisable 
to
stay with universal assertions. Probabilistic knowledge of medicine should
pragmatically be considered out of scope for any ontology of general medical 
science
- it represents way too big a task! Look at the history of medical artificial
intelligence projects in the 80's - for example, how long the Internist/QMR 
project
took for its knowledge base development for a few hundred diseases.  It would be
naive and foolhardy in the extreme for OGMS to embark on a representation of
probabilistic knowledge of all of medicine, or even a small subset.

Original comment by ksp%ihts...@gtempaccount.com on 15 Jan 2010 at 9:22

[GoogleCodeExporter]

Point taken Kent...we should strive to use any such relations only where we can 
make
universal assertions.  I think the question here is whether OGMS should provide 
such
relations.  

One of the core assumptions of OGMS is that every instance of a disease has some
disorder as its physical basis.  For the great majority of cases, there will be 
at
least some symptoms experienced by the patient localized at the disorder: pain 
in a
fractured leg, tenderness on a burned shoulder, itchiness on a rash etc.  I do 
think
there are universal facts underlying localized symptoms...and so we should 
provide
such relations...but this would not cover all symptoms in the broad sense (i.e.,
those inferred solely on the basis of probabilistic knowledge).

Original comment by albertgo...@gmail.com on 18 Jan 2010 at 2:48

[GoogleCodeExporter]

Albert, is there a particular way in which 'defined class' is being used in 
OGMS?   BTW, OBI has an interesting 
write up on this:
     http://obi-ontology.org/page/Defined_classes

I am not sure what is meant by 'I do think there are universal facts underlying 
localized symptoms' - please 
explain. 

You suggested that definition 'would not cover cover all symptoms' - this is 
good in a general sense, but 
where do we draw the line. Bill suggested pathognomonic features and this may 
be a good start. 
Pathognomonic feature (sign or symptom) means that 'when present it is (almost) 
certain to be a specific 
disease' but again it need not always be present. 

Original comment by sivaram....@gmail.com on 19 Jan 2010 at 8:01

[GoogleCodeExporter]

Sivaram,

The current terms under 'defined class' are nonuniversal, ad-hoc, or used for
epistemic reasons...this is all I mean when I say 'defined class'...we need to 
change
this b/c different people mean different things by it (e.g., a class for which 
there
are necessary and sufficient conditions), see
http://code.google.com/p/ogms/issues/detail?id=41

What I meant by saying 'I do think there are universal facts underlying 
localized
symptoms' is that the feeling of certain pain accompanying certain diseases (the
feeling of fever with influenza) and certain disorders (the pain at the 
location of a
fractured bone) should be treated as more than just accidental correlations, 
and thus
belong in OGMS...however some symptoms do not reach this level of law-like 
regularity
(to Kent's point) and are probabilistic assertions (see above).

FYI...I like the proposal for 'pathognomonic features' as given...what does 
everyone
else think?

Original comment by albertgo...@gmail.com on 19 Jan 2010 at 8:37

[GoogleCodeExporter]

I am getting confused by the thread - here is my take, which I think matches 
what
Albert means with 'pathognomonic features' and 'more than just accidental 
correlations'

As I understand, pathognomonic features are features highly specific of a 
disease. So
specific that they are present if and only if the patient has the disease.
If X is a pathognomonic feature for Disease D, and patient A presents X, then 
patient
A has disease D. Inversely, if patient A has disease D, he presents X.

I think this is useful, however it restricts to very rare cases, and I would 
want to
have the second case as well, what Albert calls 'more than just accidental 
correlations'.

Taking the Brucellosis example and associated features:
    * Fever
    * Chills
    * Excessive sweating
    * Weakness
    * Weight loss
    * Fatigue
    * Headache
    * Abdominal pain
    * Back pain
    * Loss of appetite
    * Joint pain

I understand that usually no patient will present with the whole list of 
symptoms,
but we should be able to describe "a canonical case of Brucellosis", something 
along
the line of what FMA is doing with a canonical anatomy model.

Ultimately, I would want to be able to query things like:
1. give me all diseases where patient has (may have) fever
2. if my patient has Brucellosis, what can I expect as symptoms?

I expect query 1. to return Brucellosis, influenza and probably quite a few 
others. 

Maybe one solution would be to reflect this in the name of the relation, 
something
like "has_X_symptom", where X could be canonical, usual or else?

Original comment by mcour...@gmail.com on 27 Jan 2010 at 7:21

[GoogleCodeExporter]

Following up on Melanie's last point and proposed solution...  

I think the relation building schema 'has_X_symptom' should fill in X with the
particular canonical reference intended (Google Health, WebMD, Mayo Clinic, 
Medline,
etc.)  This will generate the least controversy since there are differences 
between
what experts take to be canonical references.

So we could have:

Brucellosis has_GoogleHealth_symptom Fever

which only means that Fever is a symptom of Brucellosis according to the Google
Health reference (see https://health.google.com/health/ref/Brucellosis).  

Notice this is an absolutely true statement that has nothing to do with 
probabilistic
knowledge at all.  Google Health does make this claim about fever being a 
symptom of
brucellosis.  Also, it may be the case that some other canonical reference does 
not
make Fever a symptom of Brucellosis.

Now two questions:
(1) Are there any tools that let you build relations from a schema in this way?
(2) How to update this information as the canonical references change?

Also, it would be best if this canonical reference X had a URI and was 
represented in
an ontologically coherent way...so there could be guidelines for which Xs to 
use. 

Thoughts?

Original comment by albertgo...@gmail.com on 27 Jan 2010 at 7:51

[GoogleCodeExporter]

I don't like the idea of pushing such provenance information into the relation 
itself. 
The symptom is related to the disease and not Google Health - therefore, saying 
'has_GoogleHealth_symptom' 
makes it appear as a special kind of symptom. Google Health is only asserting 
the statement.

Melanie: I don't think one should interpret 'pathognomonic' to mean a 100% 
certain. Rather, it should be 
treated as 'almost certain' or 'most likely'. There are always differential 
diagnosis associated with most 
symptoms and signs, even for those designated as pathognomonic! 
I am not sure what is meant by 'accidental correlation' but Ken mentioned 
probabilistic knowledge and that it 
should be considered outside the scope of OGMS. My suggestion was that such 
information can be codified as 
rules, again outside the ontology. 

Is there anything like a canonical disease model? The closest I can think of in 
terms of canonical knowledge is 
information in text books of medicine and similar references. 

Original comment by sivaram....@gmail.com on 28 Jan 2010 at 3:38

[GoogleCodeExporter]

Thanks for all the explanations. 
For pathognomonic, I was looking at http://en.wikipedia.org/wiki/Pathognomonic, 
and
based and that and Bill's explanation, I understood it to mean 100% certain. 

I understand the point about probabilistic associations, indeed having fever 
doesn't
mean I have influenza and vice versa. 

Patients will rarely present with the whole set of symptoms, but when a patient
consults his physician, the physician will use the list of symptoms reported to
establish his diagnostic. This process has to be based on some kind of accepted
shared knowledge, i.e. a set of symptoms and signs which taken together allows 
the
practitioner to 1/ discard some diseases 2/ restrict the list of possibilities 
to one
or a set of diseases. 

I had assumed that associating a list of symptoms with a disease was in the 
scope of
OGMS - from the above it sounds like I was mistaken.
If this case, do you have resources to recommend where that kind of knowledge 
would
be encoded or could be added? I would be more than happy with the canonical
information from medical text books.

Original comment by mcour...@gmail.com on 28 Jan 2010 at 4:31

[GoogleCodeExporter]

Wikipedia does use the words 'beyond any doubt'. However, practically it is not 
used in a strict sense. When a 
sign is said to be pathognomonic, it means that it is 'more or less' diagnostic 
of a certain disease. A diagnosis 
(of a disease) is usually made taking into consideration the whole clinical 
picture and not from any one 
symptom or sign. Presence of a so-called pathognomonic sign (or symptom) makes 
the diagnosis almost 
certain.

If you look in the examples section (of the Wikipedia article), it goes on to 
say 'Very few of the examples here 
are pathognomonic in the true sense of the word'.   :)

I would think that associating a list of symptoms with a disease would be in 
the scope of OGMS - actually this 
may be in the scope of much narrower Sub-Domain ontologies. The problem (as I 
understand) is not whether 
it should be OGMS or not - it is a case of modeling 'sufficient' conditions 
only and not as 'necessary' or 
'necessary and sufficient' conditions. And then, some sort of rule based system 
with probabilistic knowledge 
can use the model with instance data to hypothesize a differential diagnosis 
(and this is in the realm of AI).

Original comment by sivaram....@gmail.com on 28 Jan 2010 at 5:48

[GoogleCodeExporter]

Alan has proposed that symptoms (and signs) be reinstated as defined classes 
that could either be qualities or processual entities.

ogms:Symptom is_a bfo:Quality OR bfo:Processual Entity

I believe we also need shortcut relations to tie symptom (quality) and symptom 
(processual entity) to a disease.  Obviously these relations would be 
many-to-many.  Here are my proposals:

S symptomatic_quality_of D = 
(1) S is_a (bfo:Quality AND ogms:Symptom)
(2) S inheres_in SOME (Organism AND has_part SOME ogms:Disorder)
(3) D is_a ogms:Disease

S symptomatic_process_of D =
(1) S is_a (bfo:Processual Entity AND ogms:Symptom)
(2) S part_of SOME (realizes SOME D))
(3) D is_a ogms:Disease

Original comment by albertgo...@gmail.com on 2 Nov 2010 at 3:38