Tuesday, September 18, 2018

CDISC-CT: we can do better

The song of Matt Simmons "we can do better" is at the moment very popular here in western Europe. Every time I hear it, I must think about CDISC Controlled Terminology (CDISC-CT). I will explain why.
Traditionally, CDISC-CT is published as lists, usually in Excel format, nowadays also as CDISC ODM (XML) files. The latter was a great improvement. Even more important was that each list and each item ("term") in the list received an identifier by NCI, the "NCI code".

Version management

One of the things we can however do considerably better is version management.
CDISC-CT is published every 3 months. One of the Excel files published is "Terminology Changes". In this file, we find all changes with respect to previous versions, mostly additions, but also a number of "updates", like "update codelist name", "update NCI preferred term", and even "update CDISC definition".

These are highly problematic. Any other controlled terminology in healthcare that I know has the policy to also change the identifier when a term or its definition is changed. The old identifier is then "deprecated", and users are requested to not use it anymore in new applications. For example, the newest LOINC version (2.64) contains 2920 deprecated terms (for historical reasons), and comes with a (database) table with the mapping between the old and the new code (the "mapto" table). The latest CDISC-CT (2018-06-29) lists 185 "updates", of which 44 are definition updates, and 1 is a submission value update. An example of such a definition update is the submission term "DXA SCAN" (NCI code C48789) in the "METHOD" codelist:

 As as well the "submission value" as the "NCI code" remain the same, this is highly problematic, because the meaning of the value or code then depends on the version of the codelist. In the earlier versions (before 2018-06-29) the code/term means ANY technique for scanning bone density, whereas in the new version it is redefined to such a technique where X-rays at different energy levels are used to ANY anatomical location. That this change was necessary is probably due to a problem with quality management, which might be a consequence of the very short publication cycles (every quarter) of CDISC-CT. In this case, I presume that the wording "bone" in the original definition was the problem as it essentially belongs to "SPEC" (specimen), not to "method".

So, we now have the problem that the same submission value with the same NCI code is used for i.m.o. two completely different things. Fortunately, the Define-XML team recognized this problem and will as of version 2.1 allow to indicate which CDISC-CT version was used for each individual variable/domain. Whether reviewers will use this information to distinguish between "measuring bone density" and "measuring density by X-Rays of ANY anatomical location"? They probably won't.
It would have been much much better to deprecate "DXA SCAN" and "C48789" and generate a new term and code for the X-Ray case of any anatomical location.

Post-coordinated versus Pre-coordinated

CDISC-CT is "post-coordinated", meaning that concepts are kept broad and separate and selected and joined inthe process of searching. The history (within CDISC) behind this is that at sites, either no coding was used (terms were just written on paper), or each site had different coding systems, e.g. for lab terms. With post-coordination, one can then still categorize and map such captured information to a simple term and then combine it in a useful way with other post-coordinated terms to come to meaningful data entry.

It is only in the last 10 years that healthcare is starting using pre-coordinated systems, such as LOINC (for tests), SNOMED-CT, ATC (for active ingredients), and even ICD-10 (for diagnoses). Pre-coordination is very useful, also in clinical research, as it e.g. allows to define exactly which tests need to be performed already in the protocol. This is still very rarely done however. In protocols, we still often read e.g. "measure glucose in urine", leading (at submission time) to a multitude of different glucose tests (qualitative, quantitative, ordinal), with different test methods and having different units, making results often incomparable, even between sites.

If possible, providing the pre-coordinated code for a test or property is always better. It avoids the interpretation and translation / categorization step. Even when the pre-coordinated code is known, CDISC standards still require us to provide the post-coordinated terms too. If a laboratory apparatus provides us the LOINC code, why should we then still try to translate that into a LBTESTCD, LBTEST, LBSPEC, LBMETHOD? This is error prone. So, when LBLOINC can be provided (without derivation), it should be allowed (or even mandated) to discard variable values for LBTESTCD, LBTEST, LBSPEC, LBMETHOD. Only LBLOINC is then identifying. However, there is still a lot of resistance against LOINC within the CDISC-CT and SDTM teams: it was the FDA who recently mandated the use of LOINC coding for lab tests (when the LOINC code is available), not CDISC.

Semantic versus pragmatic CT

Another problem comes with lists being incomplete or overcomplete for practical purposes. A typical example is "OTHER" and "MULTIPLE". In the "RACE" codelist (C74457), they are absent, although they are valid values for "RACE" in the "Demographics" SDTM domain. The (further correct) argument of the CDISC-CT has been that "other" and "multiple" are not races. But how does a computer program know this? If it compares a CDISC submission against the published codelist, it will not find "OTHER" nor "MULTIPLE" and thus needs to throw an error. In order to know that "OTHER" and "MULTIPLE" are allowed, a HUMAN must manually dig into the SDTM-IG and might ultimately find the section "Assumptions for the Demographics Domain Model" (p. 65 in SDTM-IG 3.2) and there then ultimately find the following paragraph:

As the SDTM-IG is still not machine-readable, this is highly problematic. Software Implementors of the SDTM-IG must over and over again try to find out this kind of information by digging into the text, and then program this into their software, which is tedious, error prone and often open for different interpretations.
Even then, CDISC-CT is not always consequent in this matter. For example, in the codelist "Acute Coronary Syndrome Presentation Category" (ACSPCAT, C101865) we find a submission value "OTHER":


"Other" is semantically surely NOT a category of an acute coronary syndrome presentation. It is just the text of a checkbox in the CRF.
A good number of CDISC codelists contain the term "UNKNOWN". Examples are the "Action Taken with study treatment" codelist (ACN, C66767), the "Cardiac Procedure Indication" codelist (CVPRCIND, C101859), the "Coronary Vessel Disease Extent" codelist (CVSLDEXT, C17998), and even the "Ethnic group" codelist (ETHNIC, C66790).
"Unknown" is surely semantically NOT an ethnic group, it is again the text on a checkbox in a CRF.

So, depending on the codelist (or CT subteam?), it looks as different criteria are used to determine whether a term may belong to that list or not. It thus seems that there is a lack of systematic approach in the development and maintenance of the CDISC codelists. This probably is due to the history of CDISC-CT: in some cases, the lists were developed by collecting every possible value that was found as a checkbox on CRFs, others were developed using a semantic approach. 

How do other controlled terminologies handle this? ICD-10 for example follows a pragmatic approach: almost every group/class in ICD-10 also has a term "other XYZ". For example, for "cerebrovascular diseases", we find as well "other nontraumatic intracranial hemorrhage" (I62) as well as "other cerebral infarction" (I63.8). So, ICD-10 does not seem to have any problem with "other" not being a semantically valid disease.

One can have discussions for many many hours about principles of developing controlled terminology, but there is an easy practical solution for this: mark those terms that are semantically belonging to the parent term (i.e. the codelist), and mark those separately that are (additionally) allowed values, and DO this in a machine-readable form, so that implementors do not need to start digging into a non-machine-readable SDTM-IG.
This might be something like: 

Or any other construct that indicates that "MULTIPLE" and "OTHER" are not real races but are also allowed submission values in the context of the SDTM-IG. One could of course further extend this with information about the version of the SDTM-IG, the applicable domain/variable, any applicable rules (preferably machine-readable) etc., but this beyond the context of the current blog entry.
For the "LBTESTCD" codelist this could look  like:

Indicating that "LBALL" is also an allowed value as a test code in the case an SDTM submission.
Please don't shoot on me on how I exactly formatted this. This is just one possibility of many. It is just about the principle of finding a compromise between keeping a controlled terminology list semantically pure (if desired) and still making it practical for its use in SDTM (or any other implementation) in a machine-readable way.

In the context of practicability, I had a customer last week asking me whether "U" ("Unknown") is an allowed value for OCCUR variables. In the SDTM-IG, the "YN" codelist ist indicated, e.g. for MHOCCUR:

referencing the "NY" codelist which, upon (manual) inspection, contains the values "Y" ("yes"), "N" ("no"), "NA" ("not applicable") and "U" ("Unknown"). So, this would mean: "Yes, U is allowed for MHOCCUR". However, manually digging into the non-machine-readable IG (p.56) also reveals:

Which would mean "No, U is not allowed for MHOCCUR".
However, we are not done yet! Further digging also reveals (on p. 40):

Essentially stating that "U" is allowed for MHOCCUR if there was a checkbox for it on the CRF.
Remind that all this regulation comes in a non-machine-readable way, making it extremely difficult to implement this in software, not to speak about differences in interpretation.
Also, this kind of rules could be embedded in a machine-readable construct as proposed above.
The same applies to the SDTM BLFL ("baseline") variables where only "Y" is allowed, this although the SDTM-IG references the "YN" codelist. In this matter, it is remarkable that the CDISC-CT has always refused to publish a "Yes Only" codelist (though requested several times). 

Lists and relations

The main problem of the current CDISC-CT however is that it is just a set of lists.
This probably is again related to that some of the CDISC-CT was originally developed for representing checkboxes on CRFs. Even now however, we still stick to lists without little or no relationships between terms. We are thus missing enormous opportunities - let me explain with a few examples.
The codelist "VSTESTCD" (vital signs test code) contains over 40 codes, from "ABSKNF" (Abdominal Skinfold Thickness) to "WSTCIR" (Waist Circumference): 

Let me remark here that, at each new version, I regenerate the "TESTCD" codelists to also contain the "TEST" value (as a "decode"), as unfortunately the CDISC-CT stopped doing this for one reason or another. In doing so, the relation between the test code and test name becomes directly visible (which I think is much more friendly for software that uses it), instead of needing to do additional joins (using the NCI code) within the software itself.
So, the VSTESTCD codelist as published by the CDISC-CT team is just a list. Of course it also contains synonyms, definitions, :

But it does not state anything about (semantic) relationships between the terms within the list itself, nor with terms in any other lists.
For example, it does not state that "systolic blood pressure" and "diastolic blood pressure" are highly related (both are cardiovascular properties), but that there is little or no relationship between "body length" and "systolic blood pressure" except that both are considered as "vital signs". Similarly, the CDISC-CT does not state (at least not in a machine-readable way) that "BMI" and the combination of "body weight" and "body height" are highly related. It would even be easily possible to express this relationship (a formula) within the CDISC-CT publications in a machine-readable way.

Another example is "ALBUMIN" in the "LBTESTCD" (Laboratory Test Code) codelist. Remark that "Albumin" is essentially not a lab test, it is the analyte that is measured in a lab test. So TESTCD has different meanings in SDTM, depending on the domain in VS it is the property measured, in LB it is the analyte that is measured. But that is essentially an SDTM problem, not a CDISC-CT problem.
Why is albumin tested in healthcare and in clinical research? What is it related to? CDISC-CT does not tell us at all. CDISC-CT does even not tell us that the "albumin test" is usually part of the "liver test panel". Another (and better) coding system for lab tests (precoordinated however), LOINC, does at least provide us this information:

And even providing the information which other lab tests belong to the same panel, e.g. telling us that a "bilirubin" test is highly related to our "albumin" test. CDISC-CT does not provide us such information, partially because "LBTESTCD" is post-coordinated and essentially is about analytes, not about individual tests (as LOINC does).
Also remark that LOINC also provides us the "usually used" corresponding units in UCUM format, a format that is still mostly ignored by CDISC-CT, and still not accepted by SDTM, although it is THE format used by electronic health record systems worldwide (CDA, FHIR,
). This is a huge problem in the era that much of our data is coming from EHRs or hospital information systems anyway. SDTM should at least accept UCUM format, besides existing CDISC-CT for units. The further development of the latter should be stopped it does not make sense anymore.

Fortunately, we now have UMLS (Unified Medical Language System). UMLS is trying to cover most of the coding and controlled terminology systems used in medicine, and to provide relations between them. Each term or code in an healthcare coding system like LOINC, SNOMED-CT and also CDISC-CT also has a code in UMLS. This allows to generate "networks of knowledge", e.g. connecting the CDISC-CT term "albumin" with the LOINC code 1951-7 (Albumin [Mass/volume] in Serum/Plasma) with the "hepatic function panel", with the organ "liver" (SNOMED-CT) and with the disease "hepatitis" (SNOMED-CT, ICD-10), etc..
UMLS even makes RESTful web services available allowing to develop applications to build such "knowledge networks". One of my students is currently developing an application (for his master project) to interactively generate and display such "knowledge networks". I will later report on this in a separate blog.
However, where possible and useful, we should not leave it to UMLS to do the work on relations between terms for us. 

Public reviews

Another problem is that the CDISC-CT team keeps publishing drafts for public review only as Excel files. Excel is not a vendor-neutral format, nor is it an open standard. This makes it difficult to do QA and impact analysis on changes relative to earlier versions. If the CDISC-CT does not do such an analysis, it should at least be made easy for the public to do so during the review period. In my personal opinion, it would also be much better to have a longer cycle (e.g. once a year) and have a longer public review period so that a broad discussion is possible prior to final publication.

What we can do better

Here are a few proposals (my opinion of course) how we can do better in the development of CDISC-CT:

  • Start better version management. For example, we could agree that the January 2019 version is used as the "base version" for the future. If something for a controlled term is changed later (such as the definition), deprecate that term and NCI code, and provide the changed term with a new "submission value" as well as a new NCI code. 
  • Improve quality management. Many of the "definition changes" are probably due to a somewhat overhasty inclusion of new requested terms. Also, there does not seem to be any "impact analysis" done on changes. Some bright people have developed methods andtools for this. These can and should also be used by the CDISC-CT team. 
  • Stop further development of codelists for which there are better alternatives, but still allow their terms to be used in submissions. The "UNIT" codelist ist the most notorious of them. Further development of the LBTESTCD codelist does not make sense in my opinion either. And why do we need our own lists of microorganisms whereas real specialists have already developed complete ontologies?
  • Allow alternative codelists that do the job better. For example, for LBTESTCD (which is not representing a test but an analyte), we may as well use the "component" part of the LOINC code. This would however require to move away from the 8-character limitation of TESTCD variables. The users could then choose between either using the "old" CDISC-CT codelist and using the "component" from LOINC for LBTESTCD. 
  • If pre-coordinated codelists can be used, do so. For example, for lab tests (but the same applies to vital signs), if the LOINC code is available, e.g. from an EHR, use it (LBLOINC, VSLOINC) and in such a case, do NOT (manually?) populate TESTCD, TEST, SPEC, METHOD, as this is redundant information and usually becomes error prone.
  • Be consequent and pragmatic. The example of "OTHER" being excluded or included in different codelists clearly demonstrates that the development principles between codelist (and CT subteams?) do differ. This is due to the history of CDISC-CT. Take a step back and rethink the develop the development principles which then need to be followed for each codelist. Best (i.m.o.) is to use a pragmatic approach with terms that "natively" semantically belong to the codelist are marked separately from terms that are allowed under certain conditions (such as "MULTIPLE", "OTHER", "UNKNOWN"). Extremely important is that this is done in a machine-readable way. One could even express such conditions in a machine-readable way within the codelist itself. And no, this is NOT the responsibility of other teams (such as the SDTM team) alone, it is ALSO the responsibility of the CDISC-CT team.
  • Where possible and useful, provide relations between terms. For example, in "vital signs", there is a clear relationship, even described by a formula, between "BMI" and "height" and "weight". This formula can even be incorporated in the CT itself. Common "parent" terms can also be added, e.g. providing the relation between "systolic blood pressure" and "diastolic blood pressure". We should not leave it to UMLS to define relationships between our terms. 
  • In order to improve the quality of CDISC-CT, make the release cycles longer again. Twice a year should be more than sufficient. Do not include new requested terms that were not well quality-reviewed and for which no impact analyses was perform. Publish public review packages in a modern, vendor-neutral format, so that future implementors can do an impact analysis on their systems. Also provide more time for such reviews. If a new term is heavily discussed, retard its publication until consensus (also with the broader community) is reached, or do not add it at all.