Terminologies and ontologies: concrete application examples
Although few people are aware of them, terminologies and ontologies play a major role in fields such as medicine.
(Bild: miguelglxz/Shutterstock.com)
Terminologies and ontologies can improve healthcare, for example in drug therapy safety, but also in other use cases.
The use cases for terminologies and ontologies in medicine [1] are not always trivial, but sometimes require quite complex algorithms. One of the biggest challenges is the processing of continuous or free text, which is still widely used in medicine.
This involves mapping phrases, sentences or entire sections with the corresponding concepts of a terminology. Often referred to in specialist literature as concept tagging or annotation, algorithms from the field of Natural Language Processing (NLP) are required for this.
Placeholders (*) and inheritance attributes(< (children of), << (element itself and children of), >> (element itself and parents of), > (parents of)) can be used and the complete syntax in the current version also includes the mapping of nested expressions, cardinalities and negations. A concept can be supplemented with a comment or label using two pipes.
In its simplest form, the expression <<* finds or expands to the total set of all "SNOMED CT" concepts. More practical would be this expression: <36989005|Mumps| which returns all (20) specializations of mumps.
Attribution can now be used to restrict such purely hierarchical queries based on the concept model. For example, all fractures that affect the femur can be represented with this syntax:
<125605004|fracture| : 363698007|where| =<71341001|on the femur|
In the case of SNOMED CT, this query currently returns 132 terms. If you replace the fracture (125605004) with an"*", you would get all 237 terms that have a topographical reference to the femur.
Interestingly, you can also invert this query and ask: Which bones can break? To do this, the focus term is simply used as the target term (note: use inheritance!) and the role is inverted by prefixing it with "R":
<* : R363698007 =<<125605004
There are almost 800 terms that satisfy this expression, i.e. structures in the human body that can break. A tool for trying out expressions can be found at https://browser.ihtsdotools.org/ [2]
Especially in German and in the field of medicine, there is a lot to master. Honeck et al. have summarized this very succinctly:
"Not only is the German language known for excessive single-word nominal compounding, but also its medical sublanguage, in particular, is characterized by a mix of Latin and Greek roots."
It is precisely at the NLP interface that potential synergies between the various AI approaches are now emerging for the first time, as the task of concept tagging can be significantly improved with large AI models (LLMs) and their methods. In particular, the processing of modalities, i.e. those elements of human language that allow assumptions to be made and hypotheses to be formulated, is now reliably possible for the first time. In future, rule-based NLP algorithms will only have an independent significance in marginal areas, for example when it comes to differentiated syntactic analyses.
"Clinical decision support system
When implementing a Clinical Decision Support System (CDSS), the aim is to provide assistance to a user, for example by drawing conclusions about diagnoses from symptoms or searching for similar patients, which can then be used as an example to plan and decide on a therapy. This requires special algorithms in the background, including so-called inference algorithms. In this article, the aspect of queries from ontologies will be explained, as there is still a clear advantage here compared to machine learning-based approaches.
The "expression constraint language" (ECL) was developed specifically for querying terms from SNOMED CT and analogously structured ontologies (SNOMED ECL). This allows formal expressions D⊆∃R.C to be represented and executed in the form focus term: attribute=target term.
Strengths compared to other approaches
The strengths compared to other approaches lie primarily in the extremely clear and efficient way in which certain facts can be defined and the traceability of the results.
In the case of "major heart surgery on children", for example, a CDSS can provide advance advice for an interdisciplinary board in order to minimize potential risks. Or, in the case of "infections in the abdomen", initiate a specific process to evaluate the cause. But how do you define "major heart surgery" and "abdominal infections"? Often simply by listing them.
First example: diagnosis appropriate?
With the help of ontologies and ECL, a formal definition can now be found and applied. A concrete example: We want to check whether the gender of a patient matches their diagnosis. To do this, we define the diseases that can only occur in male patients(<<*:363698007=<<10052007) and include over 1300 terms with this simple ECL expression. It should be clear that a list will hardly be complete - nor efficient. Try to elicit the complete list or at least a list with more than 100 entries of "male diagnoses" from a current LLM!
(Image: Perplexity.ai)
Drug and therapy safety
In any ageing population, patient care is characterized by multimorbidity and polypharmacy. Every second elderly patient suffers from more than four chronic diagnoses and takes more than five medications on a permanent basis. Often there are even significantly more medications, as many active ingredients have side effects that compensate for other active ingredients. Patients with arthritis are often prescribed painkillers such as diclofenac, but these can cause stomach problems, particularly heartburn. On the other hand, active ingredients such as omeprazole and pantoprazole are used, which in turn can promote osteoporosis. A chain that can be continued indefinitely.
Excerpt from the semantic definition of the diagnosis "hepatitis C". It includes active substances that can be used to treat the diagnosis, but also those that should only be used with caution or may not be used at all.
(Image: Wingert-Nomenklatur, André Sander)
The specific pharmaceutical knowledge represented in an ontology typically includes indications, contraindications (contraindications), interactions (interactions), cross-allergies and adverse effects (side effects). A very simple example is shown in the figure for the excerpt from the semantic definition of the diagnosis, whereby it should be noted that each of the "outer" boxes has its own semantic description. The end result is a huge network of content-related connections, which is ultimately filtered using the data of a specific patient. This makes it possible to analyze even very complex situations, such as patients with many diagnoses and many medications.
Such automatic analyses are also particularly important for medications that are used in different specialist areas. For example, certain active ingredients used to treat depression interact with those used to treat athlete's foot. In the typical day-to-day treatment of a GP, something like this can easily be overlooked or the doctor does not have all the information. If, in future, all data is stored centrally in an electronic patient file, then there is at least a theoretical possibility of implementing efficient AMTS testing throughout the entire healthcare system.
However, many active ingredients not only have drastic side effects, but also interact with each other. The simultaneous administration of simvastatin (an active ingredient that lowers cholesterol levels) and clarithromycin (used to treat respiratory infections) can lead to the dissolution of the striated muscles (rhabdomyolysis) and consequently to kidney failure. This is a life-threatening situation. Patients may also have intolerances, so the dose must be adapted to the patient's condition and age. The same applies to the duration of use and even the time at which a medication is taken can be decisive. So how can a doctor prescribe a medication safely? What support can he expect?
With ontologies that contain pharmaceutical knowledge and the "CDSS tools", an efficient tool can be implemented that addresses this problem – which is referred to as "drug and therapy safety (AMTS)" in everyday clinical practice.
(Image: Wingert-Nomenklatur, André Sander)
Standardization / Semantic interoperability
Terminologies and ontologies can not only generate knowledge, but also transport it. By mapping a textual description to a terminological term, we leave the level of human language and ultimately enter abstract mathematics or algorithms. A term can have many different narrative descriptions and is described in terms of content (semantically) via ontologies. A certain "understanding" of a term can thus be generated at the level of algorithms. The second part of the article series described [4] how terms are derived from human language and the effort involved, but there is also an effort to communicate terms from the outset instead of free text.
Instead of the word "myocardial infarction", for example, the term "I21.9" (ICD-10) or "22298006" (SNOMED CT) is conveyed. With suitable tools, such as terminology servers, these terms can be processed directly: They can be resolved backwards. This means that a display text is generated for a specific language. These can be related to each other and clarify, for example, whether a term belongs to a certain category or not. This is known as semantic interoperability. In a nutshell: This is a form of communication in which the sender and receiver not only understand each other syntactically, but in which the receiver is in principle able to understand the content of the communication.
Meta-level and description of the data
It is important to distinguish between two levels at which terminologies are used to achieve semantic interoperability. First, there is the meta-level, which describes the data descriptively. And then the data itself must be described. In particular, the content of a table with the term 419076005 (SNOMED CT for allergy) could be described at the meta level.
A system can use this to conclude that, in principle, these are diagnoses and that an allergen must exist. A line in this table can now contain, for example, 91936005 (SNOMED CT for penicillin allergy), which on the one hand defines the allergen (penicillin), but can also be validated by asking a terminology whether a penicillin allergy is an allergy. As I said, on a purely conceptual, mathematical level! However, this also works prospectively: if the term "91936005" is received, terminology servers can be used to check in which table this term must be stored.
(Image: Simplifier.net [5])
In Germany, semantic interoperability is seen as a central focus of digitalization in the healthcare sector. This keyword can therefore also be found in numerous laws. Specifications should be defined as "semantically interoperable", which is why terminology such as SNOMED CT can also be found there. The specifications themselves are usually implemented in the HL7 FHIR format.
(Image: Simplifier.net [6])
Despite all the successes of modern machine learning (ML) approaches, we should not forget tried and tested tools. All technologies have their weaknesses, and the challenge is to bring the respective strengths to the fore by skillfully combining them. Terminologies and ontologies are complex to create and use, but they give us a promising opportunity to further improve ML approaches.
Newer approaches are investigating the possibility of "enriching" data for machine learning in advance with the help of ontologies and even making it comprehensible (keyword "prior knowledge"). It should also be possible to create new ontologies more efficiently with the help of ML models. Rule-based systems are an important backbone for semantic interoperability and ensure the transparent quality that is so important for medical devices - ultimately, we want to be treated by doctors who can rely on their software.
(mma [7])
Don't miss any news – follow us on Facebook [8], LinkedIn [9] or Mastodon [10].
This article was originally published in German [11]. It was translated with technical assistance and editorially reviewed before publication.
URL dieses Artikels:
https://www.heise.de/-9805547
Links in diesem Artikel:
[1] https://www.heise.de/hintergrund/Ontologien-Terminologien-Wie-sich-Sprache-in-der-Medizin-formalisieren-laesst-9794852.html?from-en=1
[2] https://browser.ihtsdotools.org/
[3] https://browser.ihtsdotools.org/
[4] https://www.heise.de/hintergrund/Ontologien-in-der-Medizin-Struktur-und-Erstellung-9797489.html?from-en=1
[5] https://simplifier.net/erezept/kbvprerpmedicationingredient
[6] https://simplifier.net/im1x0/kbvprmiovaccinationrecordprime
[7] mailto:mma@heise.de
[8] https://www.facebook.com/heiseonlineEnglish
[9] https://www.linkedin.com/company/104691972
[10] https://social.heise.de/@heiseonlineenglish
[11] https://www.heise.de/hintergrund/Terminologien-und-Ontologien-Konkrete-Anwendungsbeispiele-9803450.html
Copyright © 2024 Heise Medien