Praxeon Blog

Human to Ontology Translation

Ontologies are formal computer scientific representations of knowledge. An ontology models the hierarchical (parent/child) relationships between concepts, and the cross-linking relationships between these concepts. For example, ontologies such as the FDA drugs database, MeSH, the NCI Thesaurus, and SNOMED can tell you that 'bupropion' is an aminoketone phenylethylamine derivative, it is an antidepressant, and it is an FDA approved drug. Therefore, once a computer receives some input and identifies the 'bupropion' concept in an ontology, there are many useful functions it can perform and inferences that it can make.

However, ontology designers (humans) are generally NOT attempting to help the computer interpret the wild and wolly free-text input that it receives from the real world. Even when a computer is talking to another computer, they may be using different encoding schemes (different ontologies). When talking to a human, the situation is even more complex because no one has even been able to get a human to adhere to a single coding scheme; we prefer to use language the way we have been using it all our lives.

So people designing and building medical information systems are left with an important problem. Our 'semantic fingerprinting' engine has been designed and developed to solve exactly this problem : identifying ontological concepts in real-world free-text human input. Other posts (Introducing Document DNA, Builts-in Synonyms) have discussed how this technology works. I'd like to take the remainder of this post to describe a couple of practical applications.

CCR Merging

The Continuity of Care Record is a specification developed for exchanging patient health information among providers. The idea is that as a patient moves from provider to provider, their CCR moves seamlessly with them. Each provider adds new information about new diagnoses, tests, drugs prescribed, elements of family history, etc. The meat of a CCR is these informational records. Each record is composed of a 'Text' name (the human readable name), and a 'Code', which identifies the record in the coding scheme (the ontology). You can see immediately what the problem is going to be with exchanging CCRs; there are many different coding schemes, with varying levels of completeness in the areas of drugs, diseases, procedures, signs and symptoms, etc. Suppose care provider A sends a CCR to B, who sends it to C, who sends it back to A. Suppose that B and C use different coding schemes than A for at least some of the information. How is A going to be able to tell which records in the CCR have changed? The Text and Codes may have changed, yet represent the same information.

The semantic fingerprint provides a robust way to compare the Text of two fields, and determine whether they are the same concept, unrelated concepts, or closely related concepts. In the first case, even though the Codes may be different, we can be sure that both CCRs are talking about the same thing, and choose whichever code we prefer. In the second case, we can be sure that the records are different. The semantic fingerprint can even help with the third case. Suppose a record goes out with the diagnosis of 'multiple sclerosis' and comes back with 'neuromyelitis optica'. In some ontologies, neuromyelitis optica is a child of multiple sclerosis. In other ontologies, it is a related disorder but not a child. We can prompt a physician to examine other information in the CCR, such as notes, to help disambiguate.

In any case, by changing the representation of the CCR from Text and Code fields to the semantic fingerprint, we can quickly identify the unchanged records and the new records, and we have a powerful tool to help disambiguate the records whose status is unclear.

Code Conversion

When providers standardize on different ontologies, a difficult translation problem arises. While each one of them has chosen an ontology to use internally, in order to communicate with each other they must be able to translate into other coding schemes.

Rather than developing a translator for each foreign coding scheme and trying to maintain it in the face of ambiguity and constant change, a provider can first translate to a semantic fingerprint (or use the semantic fingerprint as their native representation). Each bit in a semantic fingerprint can provide the code or codes for any of the source ontologies that comprise the semantic fingerprint model. Again, this capability is enabled by relying on the rigorous and extensive vocabulary of medicine to unify and segregate concepts from multiple ontologies based on their synonyms.

If the destination ontology does not contain a concept (SNOMED has the 'remittent-progressive multiple sclerosis' concept but MeSH does not; the FDA drug database contains 'AMBRISENTAN' but SNOMED does not), the system can either choose a more general concept that is available in the destination ontology ('multiple sclerosis', 'endothelin receptor antagonist'), or provide the concept in the source coding scheme, or take some alternative hybrid approach.

Concept Versioning

The body of medical knowledge is being constantly updated and revised. Guidelines are changed, new drug interactions and side effects are discovered, new drugs are approved and new indications are added to existing drugs. For this reason, as well as error correction and re-organization of existing concepts, medical ontologies are constantly changing; most are updated at least monthly, often weekly. Therefore any system which is ontology-based must be constantly revised and updated.

Each semantic fingerprint is based on a specific version. The changes between versions are available through the semantic fingerprint API, and each new version consists of a curated, consistent merging of the source ontologies. So rather than having to track and manage many ontology versions, a semantic fingerprint-based system simply stores the model version along with each fingerprinted record. When the model changes, the fingerprinted records which may have been affected can be incrementally updated.

Discovering Undiscovered Knowledge

I've written several times about the value of Curbside.MD for finding answers to clinical questions involving patients.  But something I have not mentioned is Curbside's value for research - particularly for finding associations that are not readily appreciated.  An example of such a novel connection recently arose in my clinical practice with regards to victims of trauma.

 

For years, the dogma has been that hypotensive trauma victims should be resuscitated with packed red blood cells.  But recently, those of us in the emergency medicine and trauma community have been hearing about some new ways to resuscitate trauma victims - based on the U.S. military's experience in Iraq.  Supposedly, the U.S. military has been resuscitating injured soldiers with whole blood - not packed red cells.  And interestingly enough, the outcomes have been superior for trauma patients resuscitated with whole blood compared to packed red blood cells.

 

To my knowledge, this research has not been published.  But it has been presented publicly and the word of it is spreading in the EM and trauma communities.  And it's already affecting our clinical practice.  So now, we are resuscitating our trauma patients with clotting factors, like fresh frozen plasma, earlier and more aggressively than ever before.  That's because people believe that it's the coagulation factors in the whole blood that are really making the difference.

 

So this got me wondering.  Although the military has not published their experience, has there been other research that would support out changed practice of the use of clotting factors - earlier and more aggressively - in trauma patients.  And here's where Curbside.MD comes in.

 

This is a tough thing to search in either PubMed or Google or any other medical search engine.  But let's give it a crack with a naturally phrased query in Curbside.MD.  Let's start with the query, " What evidence exists to support the use of clotting factors in the resuscitation of patients with hemorrhagic shock?" Right in best hits, the second and third articles provide some insight into our current practice.  In the second article, "Are we giving enough coagulation factors during major trauma resuscitation?" the authors argue that transfusing the equivalent of whole blood early on is the preferred resuscitation strategy.  But it's the third article, regarding experimental evidence of the use of factor VII in blunt trauma, that's really interesting.  I've written about factor VII in a previous blog. But this is a novel application that might also be useful in trauma. Now that we've found the exact article we're looking for, we can use the "See Related" feature to find every article that is exactly like it.  And that's the power of Curbside.MD.  The entire title and abstract of the third article becomes the basis of the search.

 

If we hit "See Related" for the third article, we find a wealth of relevant evidence that hints at why our changed practice might be the way to go.  There are multiple articles that suggest clotting factors, in particular recombinant factor VIIa, may be beneficial in a variety of situations requiring hemorrhage control, including for abdominal trauma, cardiac surgery and orthopedic surgery.  Now we can see why the word of mouth trauma experience from Iraq make sense.  Now we can connect the dots and tell a story, based on evidence and research, that allows us to explain our clinical findings.  And this process - finding and understanding evidence in new ways - is really enabled by the naturally phrased queries of Curbside.MD.

Come See us at WebInno 13

On Monday July 9, we'll be presenting Curbside.MD at the 13th meeting of the Boston Web Innovators group.

In addition to showing www.curbside.md, well be providing some insight into a very exciting offering that we're developing for non-professionals. We are learning a lot from our collaborations with our advocacy group partners such as the Accelerated Cure Project and Montel Williams MS Foundation. As we can see from their rapid adoption of the Curbside.MD search box, they already see value in our sophisticated medical search capabilities. They are helping us figure out how we can use the power of our technology to help patients with chronic diseases to better understand their conditions. Our solution combines the following elements to present search results that are right on target, to provide active news and alerts on subjects that are important to each individual, and to gently integrate health-related concerns into social networks.

• Understand each person's personal health history, to tailor and focus the search results and news.
• Understand the target audience of each document, to present the most comprehensible information to the reader.
• Organize documents from diverse sources including medical research, online discussions, books, blogs, authoritative web sites, and patient support groups to present a single coherent picture of the current understanding of any medical topic.
• Present related information from across all the aforementioned sources so that each page is a gateway into a world of related information.
• Accomplish this within a high-quality, moderated environment to keep the environment constructive and supportive.

Stay tuned!