Modifying the HL7 Continuity of Care Document (CCD) to Improve Reporting of Results from Functional Assessment Instruments
***DRAFT***
Tom White[1], Jennie Harvell[2], Mark Tuttle[3], John Carter3
Modifying CCD Representation of Assessment Instrument Results Page 1 of 12
Abstract
A standardization process recently endorsed by NCVHS has bearing on the Functional Status section of the current CCD Ballot. On 11/06 NCVHS endorsed a “LOINC-ification” process that supports the standard coding and messaging of patient assessment instruments that includes disability assessment and functional status content. The LOINC-ification process represents the content of the patient assessment instruments in LOINC and links LOINC-coded assessment questions and answers with usefully related semantic matches. That is, it represents both the instrument itself – the questions and possible answers, and the result of applying the instrument to a given patient. The process makes use of the UMLS (Unified Medical Language System) to represent mappings – links - between particular questions and answers, and codes from terminologies such as SNOMED, and it makes use of HL7v2 to transmit instrument results. The NCVHS endorsed the use of HL7 v. 2.4 (and higher) and CDA (Clinical Document Architecture) for the exchange of patient assessments and otherstandardized functioning and disability content. This work was advanced through and endorsed by the Consolidated Health Informatics (CHI) Initiative.
The current CCD ballot proposal for Functional Status appears to be in conflict with existing psychometric theory, and it does not take advantage of the recent NCVHS recommendations. This paper will illustrate these conflicts, describe the relevance of the NCVHS endorsed LOINC-ification process, propose how the NCVHS proposal can be utilized within current CCD modeling, and propose a small modification to the CCD ballot draft to support use of the NCVHS recommendation.
At the end of this document, we summarize the favorable results of presenting this proposal to the HL7 CCD working group.
The Challenge of Creating a Canonical Value Set for all Possible Functional Status Problems
An implied goal of the Functional Status section of the CCD is to enumerate all possible functional status deficits in a Value Set. There are many standard psychometric assessments for each of the dozens of functional status domains. There are also many federally mandated forms which use measurements of functional status for payment and quality assurance purposes. Although it might be tempting to create a unique list of functional status values within each of these domains, psychometric and survey theory caution against such an effort. Both theories dictate that even minor changes to the wording of questions or the allowable response options can significantly change the meaning of the deficits being measured. Moreover, federally funded efforts to find unique codes from existing standard terminologies which adequately capture the intent of questions on existing federally mandated reporting forms, like the Centers for Medicaid and Medicare Services (CMS) mandated Nursing Home Minimum Data Set (MDS) Version 2 form, found less than 50% “exact” matches[4]. However, this same effort found over 90% “usefully related” matches (such as broader or related matches); in this context an “exact” match can be substituted for the original matching element, and “usefully related” matches are exactly that – links between instrument elements and codes that a human or computer should find useful.[1]
Further, consolidating across different instruments the different answer sets for measurements of related functional deficits often requires expensive research to properly place the answer choices along a continuum – a continuum being one way to represent diverse assessment values. For example, NIH has spent millions of dollars and several years on the PROMIS network[5], as one of the NIH Roadmap initiatives. PROMIS, the Patient Reported Outcomes Measurement Information System, is meant to create single scales for the assessment and representation of “pain”; the goals of scale are to span from 0 to “maximal pain”, apply to all age groups, and be easy to assess (using computerized adaptive testing techniques). This has required significant statistical effort using Item Response Theory to both place the various different pain scale measurements along a single continuum, and also create computerized systems for accurate patient self-report of pain along that continuum. Were HL7 to try to create canonical lists of functional status types and values, a task considerably broader in scope than the PROMIS initiative, HL7 should expect to incur similar expenses, time delays and risk.
Fortunately, recent modeling – formalization[2] - efforts for assessment instruments have avoided this problem by supporting the creation of unique codes for all elements of each standard assessment instrument, while at the same time facilitating the creation and use of semantic links – mappings - between instrument concepts and external terminology standards. This obviates the need for creating single master lists of functional status descriptors, and instead empowers users to report exactly the results they found using their preferred functional status assessment instruments in their natural form. This also empowers government and researchers to study the reliability and validity of those measurements over time, using a common framework. This supports efforts to create new, improved instruments building upon the best existing content. It also supports enhanced policy decisions (such as payment and quality policies) through the use of the best subset of available instruments and/or improved instruments. Again, both original instrument uses may be preserved, and new uses, including more productive completion of old uses, are supported by the LOINC codes, semantic matching, and standard messaging.
Modeling Content within Assessment Instruments: “Competency” as an Example
LOINC personnel have invested significant resources over the past several years to represent survey instruments[6],[7],[8] in LOINC. Since Psychometric[9] and Survey Theory[10] show that even minor changes to the wording of questions and their response sets (value lists) can change the perceived meaning of the deficit being measured, LOINC stores the full text of survey questions and all allowable answers, when available. Moreover, LOINC has modeled both entire instruments, and the sections within them, so as to support re-use of the individual questions and answer sets which may be shared across instruments. A discussion of instrument-level modeling is beyond the scope of this current discussion, but may become relevant to future versions of the CCD.
This process of LOINC-ifying assessment instruments and linking them to usefully related terms has been endorsed by NCVHS[11]. Starting in 2005, the Department of Health and Human Services funded, under the leadership of the Office of the Assistant Secretary for Planning and Evaluation (ASPE), an effort to standardize CMS’s mandated assessment instruments for long term care, including the Nursing Home Minimum Data Set (MDS). This work resulted in a white paper recommending the LOINC-ification strategy for encoding the content of assessment instruments, ensuring that content can be messaged via HL7, and facilitating efforts to identify, store, and use semantic mappings to the concepts measured by the MDS in standard terminologies.[12] The white paper also noted that, in principle, this LOINC-ification/semantic matching process is applicable to any survey content, regardless of content domain. Thus the process is one strategy for overcoming the many gaps in data standards. One gap is the normalization shortfall – unidentified redundancy across instruments; another gap is the presence of instrument elements that lack a representation in standard terminologies. After meetings with relevant federal agencies, HL7 workgroups, standards development organizations, and representatives of SNOMED, the proposal was unanimously endorsed through the Consolidated Health Informatics (CHI) Initiative[13] and presented to NCVHS in October, 2006. NCVHS endorsed this proposal, conditioned upon the resolution of a few issues. The conditions applicable to the exchange of patient assessment content have been largely resolved.
This LOINC-ification/semantic matching process has been approved by CHI and NCVHS, applied to several large instruments used to collect functional status information, and it has been shown to be compatible with HL7 RIM[14] and CCD modeling. Therefore, since it can apply to assessment instruments in general, we recommend that HL7 adopt this approach to representing and messaging functional status content instead of trying to create canonical value lists.
The following example shows how the LOINC-ification process models, encodes, messages, and supports semantic interoperability for one of the many measurements of functional status within the MDSv2. Figure 1 is a screenshot of question B4 from the MDS, which assesses “competency.”
Figure 1. Question B4 from the paper MDSv2 form.
Figure 2. LOINC 2.17 (RELMA 3.17) representation of MDSv2 question B4.[15]
As shown[16] in Figure 2, LOINC encodes all of the content and contextual information from question B4. LOINC creates a code for the entire instrument (45981-8), and all relevant sections within the instrument (e.g. 45987-5 – Cognitive Patterns Section). It also creates unique codes for each unique question within the instrument (e.g. 45490-0 refers to the MDS Question B4). Moreover, as seen in the bottom of the Figure, LOINC stores information about all allowable answers to question B4, the order (seq#) in which they appear in the answer list, the value which the source system associates with each answer (Code), and optional spaces for Global IDs (e.g. concept codes for those answers, if relevant). In addition, LOINC stores contextual information determined to be relevant to the meaning of the items, such as consistency checks, relevance equations, and other “help” content.
The ASPE-led MDS work found that there can be zero to many semantic matches between instrument “answers” and elements of existing terminologies. For example, Figure 3 shows that there are three SNOMED codes which are usefully related to the answer 2 (MODERATELY IMPAIRED) for question B4.
Figure 3. There are three SNOMED codes usefully related to answer 2 (Moderately Impaired) for Question B4.[17]
1. MSH|^~\&| *
2. PID| *
3. OBR||||45981-8^MDS FULL ASSESSMENT FORM^LN| *
4. OBX||CE|45490-0^MAKES DECISIONS REGARDING TASKS OF DAILY LIFE^LN^B4^Ability to make decisions regarding daily life^MDS||2^MODERATELY IMPAIRED-decisions poor, cues/supervision required^MDS^F-90157^Difficulty using decision-making strategies (finding)^SNM| *
5. [additional OBX segments – one per question / measurement]
Figure 4. Sample HL7 message to transmit answer 2 to MDSv2 question B4.
Further work with LOINC and HL7 developers concluded that this content could be messaged within HL7 version 2.5 within the OBX-3 and OBX-5 message segments, as shown in the Figure 4. In Figure 4, Line 3 is the optional OBR message segment which names the instrument that is the source of these questions. Line 4 shows the OBX segment which indicates that answer code 2, Moderately Impaired, was selected for MDS question B4. The blue section is OBX-3, which uniquely names the question as 45490-0, and also indicates that the associated name within MDS is B4. The purple section is OBX-5, which transmits the value for this observation, indicating that the value is 2, per MDS’s coding system, and that the best usefully related semantic match for that value is SNOMED code F-90157. Additional name/value pairs for questions would be transmitted in subsequent OBX segments.
The ASPE project also found a way to use the UMLS to overcome limitations in HL7’s ability to transmit a repertoire of semantic relationships. For example, one limitation of the HL7 2.5 messaging is that only a single associated code can be transmitted in the OBX-3 and OBX-5 segments. For OBX-3, that associated code must be the variable name of the source system (e.g. B4) so that when HL7 messages are received, the source system knows how and where to store the values sent in OBX-5. OBX-5, on the other hand, transmits the value expected by the source system, plus one other usefully related code. As Figure 3 illustrated, there can be many usefully related codes.
The UMLS Metathesaurus[18] table MRSMAP[19] supports the representation of many-to-many relationships among concepts and coding systems. Figure 5 shows how the UMLS MRSMAP table formally represents the assertion that SNOMED codes F-90120, F-90156, and F-90157 (see Figure 3 for their Text values) are similar to answer 2 to MDS question B4. One barrier, which will be eliminated by the end of 2006, is the lack of a coding system to uniquely name instrument “answers.” LOINC will be enhanced to provide LOINC_ANSWER_PART codes, called LA codes, to uniquely represent case-insensitive answer strings. These LA codes will not represent concepts. Rather, the combination of an LA code for an answer and the LOINC code for the question represent mappable concepts. Thus, the concept implied by answer 2 to MDS question B4, which is roughly equivalent to “moderately impaired ability to make decisions about daily life”, is similar to SNOMED concepts like “difficulty using decision making strategies (finding)”. The FROMEXPR uses a Boolean expression to indicate that the “from” concept is the conjunction of the LOINC code for question B4, and the LOINC_ANSWER_PART (LA) code for the string “MODERATELY IMPAIRED-decisions poor; cues/supervision required”. The TOEXPR is the “usefully related” SNOMED code.
Figure 5. Fragment of planned UMLS MRSMAP submission to codify mapping between particular answers to MDS questions and usefully related SNOMED codes.
More complex mappings are expected, and can be supported by the UMLS’ MRSMAP schema. For example, if one looks at Figure 3 more closely, one will see that SNOMED codes F-90120 and F-90157 are also usefully related to answers 1 (Modified Independence) and 3 (Severely Impaired), respectively. None of the single SNOMED codes listed captures the full intent of the question and the concepts an answer of 2 (Moderately Impaired) represent. Rather, the most exact mappings are likely to be to a Boolean combination of post-coordinated SNOMED terms. Work on identifying such best matches is underway. In addition, broader and partial matches have already been identified, and the CHI Disability Workgroup expressed interest in having those partial matches represented within UMLS so that all findings related to particular domains, like mental status and capacity, could be easily extracted and publicly accessible.
Storing such semantic relationships within UMLS enhances the potential for reuse of survey and administrative data – both the instruments themselves and the results of applying those instruments to patients. Nursing homes are required to complete MDS forms for all patients at least quarterly, and those data are used by CMS for payment decisions and quality reporting. There is considerable clinical, safety, and functional status content within the MDS, that could be usefully included in a Continuity of Care Document. By LOINC-ifying the MDS and representing semantic mappings of MDS answers to SNOMED, MDS results could be re-used for multiple purposes including exchanging standardized functional status summary information as patients transition across health care settings and providers, as well as clinical, quality, safety, and other management purposes.