Discussion of allergic disease model with Jeremy Rogers

Hi Jeremy. With regard to the 2nd option below, the problem is that the range for due to does not include substance and for the first option I would suspect there might be some objection to role grouping a due to relationship with a causative agent relationship. I will however test the 2nd option and compare the inferences with my proposed model.

In the case in hand, therefore, I would consider e.g.

Allergic contact palmar dermatitis caused by colophony (disorder)
==
EITHER
64572001|Disease|:
{
42752001|Due to (attribute)|=418579000|Allergic reaction caused by colophony (disorder)|,
246075003|Causative agent|=255840003|Colophony|
},
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}
OR
64572001|Disease|:
{
42752001|Due to (attribute)|=418579000|Allergic reaction caused by colophony (disorder)|,
42752001|Due to (attribute)|=255840003|Colophony|
},
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}

The above example is interesting because I have also proposed a model specifically for allergic contact dermatitis that I presented during the ECE meeting in Bratislava. The above is under modeled in terms of capturing the contact dermatitis part. My proposed model could be tweaked to fit your model. I would like to add your discussion to a future ECE call in order to obtain others feedback on your comments. I will add our e-mail discourse to the confluence page.





Bruce


Bruce J. Goldberg, M.D., Ph.D.
Physician director, S.C.P.M.G. Regional Allergy-Immunology Lab
Kaiser Permanente HealthConnect
Convergent Medical Terminology
Phone:
818-392-7292 (lab)
510-368-0270 (cell)

From: "ROGERS, Jeremy (NHS DIGITAL)" <jeremy.rogers@nhs.net>
Date: Wednesday, October 25, 2017 at 6:17 AM
To: "Bruce J. Goldberg" <Bruce.J.Goldberg@kp.org>
Subject: RE: Allergy editorial guidelines

Caution: This email came from outside Kaiser Permanente. Do not open attachments or click on links if you do not recognize the sender.

Hi Bruce – apols for incomplete and possibly slightly terse email yesterday; I got dragged away by the kids and so had to rapidly ‘cut and run’..

after allergic sensitization relationship is being removed from the model

So (still playing catch-up) … the modelling of both ‘allergic disposition to X’ and ‘pseudoallergic disposition to X’, is being revised so that both can follow a common modelling pattern and so automatically co-classify under the common grouper ‘hypersensitivity to X’? As you say, modelling the allergy half of that with an ‘after sensitization to X’ would require the pseudoallergy half to be modelled using some horrendous kind of negation, such as ‘not-after sensitization’ or ‘after non-sensitisation’, in order to follow the same modelling pattern and thus be co-classified.

I’d therefore support moving the still unavoidable negation into a negated pair of primitives such as allergic/pseudoallergic process; that seems to be the best mechanism available to us with which to represent negation. The modelling change this permits is also, I think, more consistent with ‘the words’ in as much as the modelling is now in relation only to the essence of what the future reaction will always be, rather than to how you may have developed that propensity. If we ever wanted to re-forge a modelled link between the notion of ‘propensity to allergic reaction to X’ and the prior ‘sensitization to X’, it could be done in future as a GCI – though not without introducing the potential for some degree of ‘shared variable ‘ conflict nightmare, in which the expressivity of the DL can’t represent that ‘propensity to allergic reaction to X after sensitisation to Y’ implies that X is a Y. And even if it could, the classifier would I think explode trying to enforce such a rule.

[The residual question of what should the link be between allergic propensity to X and sensitization to X did also make me wander/wonder more generally about the question of how SNOMED should support recording sensitisation and desensitisation (whether or not therapeutically achieved). At the moment, all the ‘desensitisation/hyposensitisation’ codes seem to assume that established hypersensitivity is only ever lost as a result of active desensitisation therapy. But maybe that’s an area to sort out in the future...]
modeling of allergic diseases … a nested construct would be ideal. Until this is available…

Its already possible to achieve the nested representation now, but only where the nested ‘allergic reaction to X’ entity subgraph can also be referenced as a precoordinated concept so that the nesting is only implicit, as in:

Allergic contact palmar dermatitis caused by colophony (disorder)
==
64572001|Disease|:
42752001|Due to (attribute)|=(281647001|Adverse reaction|:
246075003|Causative agent|=255840003|Colophony|,
370135005|Pathological process|=472964009|Allergic process|),
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}
==
64572001|Disease|:
42752001|Due to (attribute)|=418579000|Allergic reaction caused by colophony (disorder)|,
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}

Unfortunately at the moment its actually only rarely possible to do this - but only because the potential space of precoordinated concepts below 418634005|Allergic reaction caused by substance (disorder)| is currently underpopulated.

So, I would personally support a ‘Disorder due to allergic reaction to X’ model that clearly referenced a freestanding ‘allergic reaction to X’ entity as a due to subgraph:

use allergic reaction to substance as the range for the due to attribute

..but, as you say, this brings us to the next problem:

This .. might be problematic from the viewpoint of querying by causative agent

If the above pattern could be followed more universally then - in an ideal world and with a more expressive logic - all similar expressions like the above would be enabled to also autoclassify under one (but never simultaneously under both) of:

64572001|Disease|:42752001|Due to (attribute)|=255840003|Colophony|
64572001|Disease|:246075003|Causative agent|=255840003|Colophony|

…by the simple addition of one of two alternate property chain axiom possibilities:

Either 42752001|Due to (attribute)|o RoleGroup o 246075003|Causative agent| implies 42752001|Due to (attribute)|
Or 42752001|Due to (attribute)|o RoleGroup o 246075003|Causative agent| implies 246075003|Causative agent|

…neither of which is a million semantic miles away from the truism that due_to is in any case properly a semantically transitive relationship [a due to (b due to c) implies a due to c].

Note, however, that I’m pretty sure you can’t achieve classification under both groupers at the same time without putting BOTH property chains in. But this is probably an inherently logically contradictory thing to do and so I think impossible. Certainly, all the classifiers I’ve just tried it with throw an error.

Ultimately, therefore, we have to decide whether e.g. hay fever is a subtype of ‘disorder due to pollen’ or ‘disorder causative agent pollen’ since it can’t be both. This is a question that may rapidly reverse us into asking what the semantics of ‘causative agent’ ever actually was, and whether its really any different from ‘due to’ when appearing specifically with a substance or organism as the value.

Notwithstanding all the above, a general modelling principle is perhaps exposed by the above analysis: whenever seeking to replicate what would happen easily and automatically with relatively simple additional semantics and a more expressive logic, but that must necessarily happen more painfully and manually within the limits of SNOMED’s current logic, I would usually first consider manually asserting the exact same semantics that would otherwise be automatically inferred in that more expressive setting. Apart from anything else, this approach provides some degree of future proofing against the possibility of SNOMED’s own logic changing. Which, of course, it is about to do (!).

In the case in hand, therefore, I would consider e.g.

Allergic contact palmar dermatitis caused by colophony (disorder)
==
EITHER
64572001|Disease|:
{
42752001|Due to (attribute)|=418579000|Allergic reaction caused by colophony (disorder)|,
246075003|Causative agent|=255840003|Colophony|
},
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}
OR
64572001|Disease|:
{
42752001|Due to (attribute)|=418579000|Allergic reaction caused by colophony (disorder)|,
42752001|Due to (attribute)|=255840003|Colophony|
},
{
116676008|Associated morphology|=23583003|Inflammation|,
363698007|Finding site|=70887009|Skin structure of palmar area of hand (body structure)|
}

…to be each of two alternate, manually modelled candidate patterns that directly correspond most closely to what the more expressive logic would do for itself; the first pattern assumes we have first agreed that this conditions like this should not autoclassify under ‘diseases due to X’, the second not under ‘disease causative agent X’. In the expressive logic environment, the appropriate one of the two candidate property chains would have been implemented; in the less expressive current environment, we manually assert what would otherwise be inferred.

Although either of these two representational permutations looks like two independent ‘causes’ are being grouped in RG1 (especially in the second, where there are two due to relationships in one RG), this would actually be only an artefactual result of manually simulating the result of a semantically legitimate property chain: both representations would be deliberately but redundantly stating the same semantics twice, not two truly different and independently acting causes.

Of course, the residual problem with this kind of deliberately redundant representation is that authors can and will forget to add the redundant bit, without which the autoclassification won’t work. And that’s the main attraction of a more expressive logic – though note that (so far) adding in property chains is only currently being considered in the special context of the drug ontologies. The scope/requirement for some quantity of property chains to better coordinate other areas like this one is not yet being actively investigated.

Jeremy