Position Paper on the use of Negation Prepared for EHR4CR WPG2 by Prof Dipak Kalra and Julie James, UCL April 2012 Requirement The core requirement is to have a mechanism to provide safe, unambiguous machine readable “negative” statements that would be made in an EHR and that would be uploaded into a clinical data warehouse, and indeed that would be useful to query against. The following are examples of such negative statements: Patient does not have allergy to penicillin Patient has not taken paracetamol Patient does not have diabetes mellitus Finding of spleen palpable not present Diagnosis of MI not confirmed by cardiac enzymes It is well known that expressing negation clearly can be challenging; in human language – whether spoken or written – there are complexities, such as the "double negatives" which can slip in and cause potential for misunderstanding (for example: “I’m not going to cancel the appointment” – the appointment still stands, despite the negation (“not” and the negative verb “cancel”) and then a clinical example - “the patient did not have 'no protein in urine'”). In expressing information for machine processing, the challenge is even greater, since decisions may be made based on the results of that processing; denying a patient with suspected meningitis an immediate dose of benzylpenicillin because of incorrectly processed "does not have allergy to penicillin" information could have very serious consequences. The challenge too, is not just to construct a framework that is manageable for a gifted informatician, be that an information modeller or a terminologist, but to construct something that every clinical system can implement safely, and that every user can interpret clearly. In EHR4CR it is important, both for current work and even more crucially as the project moves towards the more clinically challenging developments in clinical trial data collection and pharmacovigilance, to describe the principles that will enable clear, unambiguous machine readable negative statements to be made, processed and queried. Using the Information Model? A negation attribute on a class can be used to express negative semantics for the whole class or for attributes of the class, but it is classically difficult to disambiguate which of these options applies. The following are examples of such ambiguous negation: A "negated observation of (palpitation of) heart apex beat" could mean that the beat was not palpable, or it could possibly mean that there was no abnormality found (i.e. it was normal) or it might mean that the examination itself was not performed – although this latter case could be indicated using a flavour of null – possibly NASK (“not asked” – the information has not been sought) "Type II diabetes mellitus was not observed in Patient X on 10th June 2009" expressed as the negated observation of type II diabetes mellitus with the given date, could mean that Patient X did not have diabetes mellitus on that date, or it could mean that the person who made this observation had no evidence that Patient X had diabetes mellitus, when they examined them on that date "Appendectomy not performed as planned surgery" described as a procedure, with type "planned" and a code of "appendectomy" and with a negation indicator – does this mean that an appendectomy was not performed, or that it was performed but as an emergency rather than as planned surgery? A negated administration of morphine with the intravenous route – does this mean no morphine given at all, or no intravenous morphine given (oral, intramuscular or subcutaneous morphine may have been administered)? Coded information of "duodenal ulcer" but, associated with this, some text stating "not visible on endoscopy", leaving unclear if a duodenal ulcer was looked for but not found and so has been excluded as a diagnosis, or is known to be a problem but not was visible on this occasion. The following two examples show how if both the information model (record structure) and terminology both indicate negation, it is particularly difficult to know what was truly meant and therefore what the action for that should be – should the double negation be taken to be emphasis, or should it be taken to mean the positive assertion? Or is it expected that the negative statement has been made using both alternatives (information attribute and terminology) and either one should be used but not both to give the negative information? Due to this ambiguity, some systems might decide to disregard this sort of data altogether, which in itself can be a problem. A negated observation of "ankle reflexes absent" – does this mean that the reflexes were present (with a sense that it was expected that they would be absent) or does it mean "this is a negative finding, which is confirmed by the text that the reflexes were absent"? A negated observation of "meningism absent" could possibly be understood to be emphasising negation – "there really is no evidence of meningism" – because to use a double negative for a positive assertion in this case would be extremely unlikely Negated Relationships Another method of describing negation in an information model concerns the relationships between classes - having a negation attribute on a relationship to express negative semantics between those classes. But this suffers from similar ambiguities, in particular the double negation problem, and also the risk of negating a null flavour value. Applying negation to a container relationship is even more ambiguous, leaving open to interpretation whether it applies to that node only or also to all of its children. The following are some examples of ambiguity caused by negating a relationship: A diagnosis of myocardial infarction is related to a finding of cardiac enzymes, with a negation on the relationship; does this mean that the diagnosis stands, but that the cardiac enzyme test does not provide evidence, or does it mean that the cardiac enzyme finding negates the diagnosis of myocardial infarction? A condition of anaphylaxis is noted, and has a causal relationship to an administration of an X-ray contrast medium, but that relationship is negated; this could be interpreted as the anaphylaxis was not caused by the X-ray contrast medium, but it could also be interpreted as the patient has had an administration of an X-ray contrast medium and it did not cause an anaphylaxis event A substance administration of methotrexate is documented, and is related to a diagnosis of rheumatoid arthritis by an indication/reason for relationship which is negated. Should this mean that methotrexate is not being administered for rheumatoid arthritis (but for some other condition) or is it meaning that methotrexate is contraindicated for treatment of rheumatoid arthritis in this patient? Querying Querying for negation expressed through an information model would require explicit processes, identifying negation as an attribute on a class and on a relationship, and agreeing and documenting clearly what this would mean in each combination of circumstances. Every import of EHR data into a clinical data warehouse would need to reliably interpret and map the imported data into these agreed use patterns. This would be a significant amount of work and experience shows it is likely still to be fraught with the possibility for misunderstanding. There would also need to be explicit business rules to manage situations such as double negation. This is a degree of complexity best avoided. Summary Using negation in information modelling is challenging, either within a class and its attributes or in relationships between classes; it has the potential to allow different interpretations that themselves are difficult to differentiate. This is less than ideal, and indeed the opposite objective of the basis of information modelling itself, which is to provide clarity for the domain of study. We suggest that EHR4CR does not attempt to use the information model to express negative semantics. Using Terminology (Coded Data)? Terminology expresses complete “units of thought”. For negative semantics, therefore, the unit of thought that is expressed by the terminology concept is the complete negated unit – “HIV negative” for example. This is clear and unambiguous in its meaning, as a “unit of thought” should be. And due to the nature of such expressions, a single vocabulary concept can provide quite rich expression, beyond the simple “not” or “negative” through to “absent” and “missing”. Rich terminologies in the domain, such as Snomed CT, provide support for such negative concepts. However, it is still possibly to use terminological means to express negation with those terminologies that do not support pre-co-ordinated negative concepts (e.g. ICD or MedDRA), but using an appropriate post-co-ordination syntax. This allows for the concept to be correctly described using the CD datatype. It is desirable that this be developed in conjunction with the terminology provider on a case by case basis, such that the negation concepts and post-co-ordinated expression can accurately be identified by the correct OID mechanisms, but if this is not possible, it is still to implement this process. Relationships should always have a description of their purpose, which, if they are to be machinereadable, will involve the use of terminology. Therefore, when the requirement is to express a relationship that has a negative sense (for example a “contra-indication”), it is deemed to be more acceptable to have a rich set of relationship expressions such as “is evidence for” and “is not evidence for” rather than to use a negation property (“is evidence for + negation indicator) – which is therefore a terminological approach. Querying Querying for particular pre-co-ordinated concepts from a terminology would have no effect on an existing query model which can query for one or an expanded set of concepts. Querying for post-coordinated expressions is slightly more complex, but these situations could be identified by recognising the OID for the post-co-ordinated expression and then applying the appropriate logic. Negation of Non-Coded Data? Are there any types of negation that cannot be handled by terminology (i.e. that would use nonconcept derived datatypes)? Should negation be applied to a value, a date, for example, or a MIME file or an address? Statements such as: the systolic blood pressure was not 135mmHg the date of the prescription was not 10th April 2012 the patient's address is not 10 Acacia Walk, Eastbourne are highly unlikely to be encountered in an EHR, and particularly in the sort of data that would be handled by a Clinical Data Warehouse. Therefore it would seem that all the requirement to manage expression of negation can met by the use of terminology. Conclusion It is generally accepted that where a negative sense can be appropriately communicated using a terminological expression, this is the preferred way to do it, and to date, examples of negation that cannot be managed by terminology have been not been found (note the double negative here!). Additionally, it is felt that querying for negation as expressed using coded concepts could use existing mechanisms. We therefore suggest that EHR4CR embraces a terminological approach to express negative semantics. Use of Null Flavours The “Null Flavour” is a property of the attribute that would contain a data value representation, used when a data value cannot be provided, which asserts a reason as to why a value is not provided. It is suggested that EHR4CR supports a limited set of null flavours taken from ISO 21090, as required by the business use cases, and that business rules as to how to handle these null flavours when encountered by the query process be developed. These rules are likely to vary depending on which of the key areas of the EHR4CR project is being (protocol feasibility, patient recruitment, clinical trial execution, adverse event reporting). Appendix 1 - Negation in HL7 V3 HL7 Core Principles has a special section on negation, since it can have such profound impact on the meaning of information conveyed. Core Principles states that the Act.negationIndicator has been deprecated and retired since it conflated the meaning of Act.actionNegationInd and Observation.valueNegationInd. Unfortunately CDA R2 retains this deprecated attribute, but it is clear that its use is problematic. There is Act.actionNegationInd; this may be used on any act and signifies that the negation applies to the level of detail conveyed by the attributes (and possibly associations) of the act. The interpretation of Act.actionNegationInd varies by mood; in EVN (event) mood, it indicates the event did not occur; in RQO (request) mood, it orders the action not to take place; in permission mood, it indicates that the permission is not granted. It can be used in DEF (definition) mood, to indicate that the act cannot be performed. So an actionNegationInd of "true" on a SubstanceAdministration Event with a subject and medicine X specified in the consumable participation would indicate "this subject has never been administered medicine X". If there is more detail specified, such as a SubstanceAdministration Event that had an effectiveTime of 12:05 on January 6th, 2010 and a performer of John Smith, then it would say "this subject was not administered a substance on January 6th, 2010 at 12:05 by John Smith". It does not preclude the possibility of a substance having been administered by John Smith at 12:04 or 12:06, nor at 12:05 by Jane Smith. There is also Observation.valueNegationInd, which simply indicates that the specified value was in fact not observed. This allows the explicit statement of a negative finding, such as absence of a tumour, rash, illness or other condition. The semantics are not influenced by any other attributes on the Observation. However, it is influenced by the level of detail specified in value. If the Observation.value is "severe, sudden onset rash", a negated observation only indicates that specific type of rash was not found. It says nothing about a moderate or slow onset rash. However, the RIM usage notes are clear that Observation.valueNegationind should only be used when the negated observation cannot already be conveyed by the code system used to communicate the observation value itself (i.e. by using a pre-co-ordinated or post-co-ordinated negative assertion concept). There are also negation indicators for use on relationships – specifically ActRelationship, Participation and Role. These indicate that the relationship between the two classes described by the attributes on the relationship class does not exist. It makes no assertion about the existence of the source and target classes, nor of the characteristics of those classes, merely establishing the presence or absence of the relationship with its specified properties. As a rule, Role.negationInd being true only makes sense when traversing from playing entity to scoping entity or vice versa. It is not appropriate for roles traversed to or from via participations. So, for example, a negation indicator set to “true” on a RSON (reason) relationship between a Substance Administration and a diagnosis observation would indicate that the Substance Administration was not indicated for the diagnosis, but it would make no other assertion about that Substance Administration. Note that in HL7 messages, in all cases, the negation indicator attributes default to false. This default holds even if negation indicators are omitted from a model. Negation indicators cannot be added into a new version of a model where they were not previously present, nor can they be added as local extensions because of their impact on the meaning of all other information in the model. It is therefore possible to use the HL7 negation indicators in an information model to convey negative semantics, but it is not a simple process and it is not to be undertaken lightly. It also makes querying for negative statements somewhat complex, requiring explicit query processes for negated statements. Appendix 2 - Negation in ISO EN 13606 Discussions of a similar nature to those outlined in this position paper took place during the development of 13606, noting the early work being undertaken in HL7 Terminfo at that time. After considerable discussion and consultation, and as approved through ballot: the 13660 Part 1 Reference Model for representing an EHR Extract has no information model property for negation: terminology values are the only mechanism for negation supported by the standard; those semantic relationships (between parts of an EHR) that carry a negative sense (e.g. disagrees with, declines, contraindication for) have explicit terms for those relationships defined in Part 3, and no negation property is included in the Link class.