| 85min | Review of Claude's Proposed Revised Method | Group | - Claude: For actions, "discernible" does not apply and will need to be replaced (later determined to be replaced with "topic"). Everything is blue sky - everything is very simple and is one field for ANF. For CIF, each attribute represents a single concept and makes no use of SNOMED attribute concepts.
- Claude: Let's say you have a post-coordinated expression for a procedure and you have route of administration by mouth and method X and you have it in a simple information model (IM) in an attribute called Topic. When you create the IM for CIF, you would have attributes for:
- Topic/code for the procedure
- Route of administration that is written in text
- Route of administration concept that would be used in ANF for the value and another in the IM for the method and its value would be the set of concepts you would assign in terminology (not sure if this captured correctly).
- Claude: If the IM has attributes that don't exist in the concept model or vice versa, you can't go back and forth without extending one or the other. We have to assume the set of things in the model we call attributes is the same in both.
- John: it sounds like the granularity of the representation has to be identical. The terminology model can't contain something more specific than the IM and vice versa, the same for generality (one can't be more general than the other).
- Claude: Included some reasons why pure CIF is not practical and why there can't be some unified terminology of LOINC, SNOMED, etc. SNOMED is the one that is fairly in the DL space. If you have only clinical statements with single fields, then you will have tons of models so pure ANF is not realistic either. SOLOR can fill in the gaps, similar to UMLS. We (the VA) need to provide the language for cross-terminology coordination and the tools. We can relax ANF in those areas currently not supported until the tools become available.
- Claude: Pure CIF is easy to do but it's not intuitive to clinicians. They want a drop-down that shows Head CT with Contrast, Head CT without Contrast, etc. If you don't tie the attributes for an IM to a terminology, that is a strength. The biggest failure of CIF is it's developed outside of terminology (not sure if this was captured correctly).
- Claude: We can't do this with just LEGOs. Thus, he broke it out into categories of: what, how, who, when, where, why, how much, and how long. We should forget trying to distinguish between topic and circumstances for now and focus on these instead.
- Claude: Yesterday, Sarita told us what the fields for LEGOs were. He is proposing we stay closer to the LEGO model. If we do that, the fields are:
- Discernible (equivalent to the topic)
- Timing
- Provenance
- Value
- Claude: For this exercise, he made an assumption that we don't have a unifying terminology for everything but we will use SOLOR that will tie the multiple terminologies using a more general language, like what exists in SNOMED-CT but is more general.
- Claude: He made the assumption that SOLOR supports concrete domains.
- Claude: For actions, we will have: discernible, timing, provenance, value, and mood. However, you could eliminate mood and tie it to the LEGO itself. Rather than doing this, you would have more LEGOs, such as one for phenomenon measurement, one for action request, etc. This is how we have arrived at the top 4-5 classes we've discussed with Keith. For action request, you would pick the Action Request type of LEGO vs. indicating "request" as a mood for a LEGO of Action.
- Claude: You would have an ActionRequest.discernible, one for timing, and one for provenance.
- Stephanie: How do you distinguish between the different types of LEGOs, such as a LEGO for an Action Request (is there a field/something that indicates it)? Claude: It would be through a class, not through a code. We really want to avoid using a code to indicate mood.
- Claude: He would like Kirsten to weigh-in since in some way, this goes back to the work she and Claude had previously done.
- Claude: Showed an example for ANF:
- Discernible: Administration of 2 tablets of acetaminophen 100 mg TAB PI BID for pain (terminology expression, assumes a concept model can support it and can handle concrete domains)
- Timing: RESULT INTERVAL TYPE: Continue for 2 days
- Provenance: ATTRIBUTION: type: requested, who; John Smith, role: PCP, where: some practice, when: on some date
- Provenance: ATTRIBUTION: type: recorded, who: system X, role: CPOE, where: some practice, when: on some date
- Benefits of this approach
- Already have most of this in SOLOR
- Use whatever the concept model allows you to capture, assuming it can handle concrete domains and can grab from RxNorm
- Claude: Using the same example, if concrete domains are an issue, we could use the value field of the LEGO (see his email for example).
- Kirsten: We should be able to do this but practically we cannot.
- Claude: Showed another example but for calorie-restricted diet. There is a flaw in using value because the computer can't distinguish between UOMs such as calories, salt concentration, and days.
- Claude: What happens where value is N/A, such as for "Head CT with Contrast"? Everything can be captured with the concept models. Value would be left blank. Value is really an extension scheme for ANF to support us for cross-terminologies. Kirsten: What if you only have "Head CT" but no contrast? Claude: You can't assume it's with or without it and the discernible is "Head CT." Kirsten: But there is no SNOMED concept for it. It would have one with contrast but it might not have with/without contrast for every imaging procedure. The same issue exists for laterality, such as for right arm. Claude: There is an attribute in SNOMED-CT for using substance that is included for "with contrast" (see 711232001, using substance = 424361007). You could do it that way. If they didn't define a using substance, we could define a new attribute in SOLOR and either use existing concepts for the substance or add them to SNOMED. Kirsten: What happens when we have one for without contrast? Claude: You could use "CT of Head" (303653007), which does not include anything regarding contrast. If there was a rule to query for head CTs, regardless of contrast/no contrast, this one would be used. On our side, we would also try to determine if we should query for both head CTs with nothing specified about contrast AND head CTs with or without contrast. We would probably need to go back to the SMEs to clarify what was intended. To say "Head CT without Contrast" is a problem because you would need to say substance of none.
- SNOMED appears to be incorrect because it shows this as fully defined but no way to say no substance (see 383371000119108). It includes computed tomography without contrast (399331006). Claude, Kirsten, and John all believe this is incorrect in SNOMED.
- Claude: Again, we can't use value to solve the problem because its meaning will vary based on the clinical statement, such as calories, tablets, etc. If we use parameter, the attribute has to be a concept that qualifies the value itself. Example: concept=caloric content (attribute) and a value=400 calories. The value is tied to a specific concept.
- Claude: Correction to the email he sent in which he outlined his proposal: the last example should all be for Action Requests (currently shows both action request and action performance). There are two things pulled out that can't be represented due to lack of concrete domains, which are the two items shown with CONCEPT VALUE PAIR.
- Claude: This is a proposal that he thinks make it easier. We model what we can and if we need expand, we use parameter or whatever we want to call it. He proposed we work with SNOMED to propose to add attributes missing in the concept model and add to our extension or the US extension. Kirsten: Then it's not up to the contractor to decide if we can do that or not. Claude: You may want to add to your contracts that there is an assumption that we can request to add attributes to SNOMED and to that through the formal SDO process. It is important to work with the SDOs to enhance the standards vs. doing it on our own and creating everything from scratch. Then we can advocate the addition of them within CIMI. Both CIMI and VA will need the extensions, and we can make a very good case to SNOMED.
|
| 20min | Rational for and Reaction to Claude's Proposal | Group | - Claude: Kirsten, if we do it this way, is it feasible from your perspective, knowing that we don't have to handle concrete domains right now? In this example, the ANF part is the discernable, which is the bulk of what we are doing. Kirsten: For the acetaminophen 100 mg example, we would still use an RxNorm code? Claude: Yes - we did not relax that one. SOLOR provides an expression language that can potentially cut across terminology and you would put an RxNorm code, not a SNOMED code. But, it means we need to relax the expression language to support that. Kirsten: For the discernible, you would expect a SNOMED concept for administration of acetaminophen and in the direct substanct, you would expect an RxNorm code? Claude: There is an assumption here: the discernible, which should be changed to "topic," is defined by an expression by a SOLOR expression and not a SNOMED expression. SOLOR expressions can cut across terminologies. But that may not be feasible. If we need to relax the assumption, then we can parameterize the part that resides outside of SNOMED, just as did for concrete domains. Our principle could be if you have a concrete domain using a parameter of concept and value pair, you would add another parameter for SNOMED using direct substance as the concept and where value is acetaminophen 100 mg tablet, that is RxNorm. Kirsten: The discernible wouldn't be specific to acetaminophen but generic as administration of drug or medicament PO BID for pain.
- John: We need to frame this in the form of a use case, as that is what Keith has directed. Claude: We could have a use case for:
- Dr. John Smith selects an order set and picks from his CPOE system a medication order for the administration of one tablet of the dispensable "acetaminophen 100 mg tablet, orally twice daily for pain." Note that one may not exceed a total daily dosage of 600 mg and the order should be performed over the span of two days.
- John: There is another piece which is to show that Keith won't be able to do this with his current approach.
- Claude: Stephanie had also asked if we can integrate this into the document we are working on. But the problem is he has a hands-on meeting today for KNARTs and won't be able to have it ready by Monday but will trying his best to have it ready for Tuesday.
- Claude: Why is this better than what we had before?
- Can more naturally use terminology concept models and we can use the concept models more fully for the heart of ANF.
- We don't have to artificially separate the what from the how and why. It's all part of the discernible pattern, and it's more intuitive to the terminologist and the clinician. Deloitte and Sarita would be engaged to refine it. We leverage what exists and tweat it at the concept level but not try to split it.
- It follows the culture of the LEGO approach that Keith has been proposing. If you look at phenomenon measurement, it was completely different from the other models we were proposing. For the others, we have the what, the object of the target, etc. In this case, it's much more symmetrical. In the LEGO structure, because terminology couldn't handle concrete domains, it introduced the idea of a value with a result. So LEGOs chose to extend the discernible with a value and this is where you could capture the concrete domain. We propose something similar but in the LEGO, you don't have to show what the value represents because it's known for the question you are asking. But for procedures, you can't make that assumption as it could be a number of things. Thus we generalize the LEGO value and you must provide a named value (i.e., the concept value pair). If it's truly ANF, it will be weaved into the discernible but we won't worry about this right now and we'll just separate them and we have the parameters.
- Claude: That is the rationale for this approach - consistent with LEGO apprach, less ambiguity in terms of what goes where (issue of topic vs. circumstances), and it has the most potential for alignment with where SOLOR is heading, and provides an escape hatch for the VA not to have to come up with tooling for the domains we can instead use parameter and value/concept pair.
- Claude: What are your thoughts on this proposed approach?
- John and Patrick: Like this approach.
- Joey: He is basically confused at this point. Claude: If you are confused, that is concerning. If you use concept model, you have to parameterize it. Which aspects of this do you think are unintuitive? That is the challenge we are trying to resolve. Joey: Seeing additional examples may help but he is also sick and may not be thinking clearly. Claude: When you are feeling better, take a look at the additional examples and let him know his thoughts on them and if he still doesn't think they make sense.
- John: Give clear use cases and explain why Keith's approach won't work. The split between topic and circumstance just isn't doable, in his opinion. Claude: The separation really hasn't been done at a structural level but on a clinical care pathway. FHIR has a similar issue.
- Kirsten: As much fun as it is to think about ANF, she is thinking about TSRs. However, she doesn't see an immediate showstopper. When we intersect with the formal model, we will need to add principles to define the differences between lab and imaging procedures, for example, She is concerned we may have the same type of separation issue that we've experienced with the current modeling approach. it will be interesting to see what this looks like in the TSR tempalte to see if we can use it or not. She doesn't think we'll see this in the TSR until it's formally accepted. Claude: The TSR will look very similar to what you see here, and he can provide an example hopefully early next week. The value can vary and mostly it's quantities but it could be a concept. We have to figure out how to handle it in the TSR and the principles for when you put it in the parameter. Maybe we could split off as a group and work on a TSR she is currently working on and how we would do it - SNOMED, VA, SOLOR, CIMI extensions, etc.
- Stephanie: She is still in a learning mode and really can't say yay or nay at this point. It will be helpful to see the information Claude proposed int to the format like the one for the white paper we have previously been reviewing, along with seeing more examples. Claude: The issues we will face are known to the terminology community. The real decisions are made by the terminologists as they define the clinical ontology. That is the separation of concerns we really should have and the one Keith started with but strayed away from. For ANF, the issue of discussion should be with terminology. All of the principles would be done outside of terminology and with a group who understands the IM. We are using the LEGO idea and these heated discussions should occur within the terminology group. It doesn't make the problem less dimensional on the terminology side but it entrusts it to a community who has a deep experience in this space.
|