At the heart of Clinia’s Health Navigation Infrastructure Is a Knowledge Graph

Understanding and navigating healthcare information can be daunting, even for healthcare professionals. Patients and professionals involved in the medical and healing professions need to reference specialists, facilities, institutions, quickly and efficiently in order to provide or obtain the needed care immediately. 

Clinia is a Montreal based health information provider that manages health directories for medical apps that are used by clinicians, doctors, healthcare providers and institutions. 

Data management efforts are invisible to the users of the health apps. Users can simply search on a condition, a part of anatomy, a region, specialist, or medical clinic, for example, and receive accurate references of whom they need to see and where. This personalized service is available through any device and Clinia’s services are used by medical apps across all of Canada and expanding rapidly into US regions. 

Creating, managing and keeping this information relevant and up to date requires the attention of trained lexicographers, computational linguists, taxonomists, ontologists, and information architects. 

Meet Anne-Laure Jousse, Taxonomist and Product Owner for Clinia. She is responsible for the quality of the Clinia taxonomy/ontology and is all those things. How does she do it?

In this article, we’ll explore how Clinia maintains the vast amount of data that facilitate the retrieval of vital healthcare information across North America.

ALJ: My background in computational lexicography and fifteen years of experience as a lexicographer have fueled my passion for modeling lexicon, especially semantic relationships between words. More recently, I've shifted my focus to natural language processing, with a particular interest in semantic text analysis and information extraction.

Working on a taxonomy feels like a natural extension of my lexicography background. In both cases, the goal is to describe concepts, explore the different ways they can be expressed, and articulate their relationships to other concepts. Fortunately, I work with talented colleagues: Camille Demers and formerly Isabelle Bastien, who bring expertise in information sciences (and even nursing), making our team well-rounded and versatile. At Clinia, our work as taxonomists not only involves maintaining the taxonomy, but also collaborating on various NLP projects related to the search engine.

ALJ: Clinia's health navigation infrastructure powers millions of health journeys across Canada. On the patient side, this includes self-navigation experiences using the Clinia Search API, embedded into our clients telehealth and virtual care applications and portals to help their members and patients find and access the right health resources inside their ecosystem and network. On the clinical provider side, the Clinia Directory adds a prebuilt UI on top of our health navigation infrastructure, to empower over 400 care providers to find, access, and recommend resources from within their own trusted portal.

ALJ: Healthcare providers access the directory app directly and can customize it with an API. Third-party applications use the search engine via an API and create their own interface for their clients.

ALJ: Our taxonomy undergoes both manual and automatic updates on a regular basis. Our daily tasks involve making manual additions, deletions, and modifications. For quick queries, we often use the Sparql endpoint in Graphologi. However, when adding information in batches, we import data through the API. The API allows us as well to search the history for specific modifications, automate much complex queries, and integrate them into global processes.

As a team, we require an efficient collaborative tool to maintain the taxonomy. Graphologi is a valuable asset as it allows us to edit the taxonomy simultaneously and provides collaboration features for discussing content within the tool. For example, suggestions for concept additions or modifications can be made through the suggestion feature and reviewed by peers at a later time. In the same vein, we recently began using SKOS collections to tag concepts that require further brainstorming. 

On the first day of each month, we conduct a monthly QA during which we review taxonomists’ suggestions and ensure the adequacy between our labels and associated language tags. In addition to Graphologi's quality check that detects inconsistencies, we also run a custom taxonomy validation script that tracks missing or excessively long labels, inconsistent use of classes in concept schemes, and other issues.

ALJ: The taxonomy is updated on a daily basis. We are constantly adding terms, concepts, relationships, definitions, various identifiers, and so on. An up-to-date version of the taxonomy is integrated into the search algorithm about every two weeks.

ALJ: Our taxonomy is multilingual. For now, we mainly cover English and French, as well as certain regional variations of these languages (Canadian/US English vs. British English, Quebec French vs. Hexagonal French). We will certainly be adding other languages in the future.

ALJ: We have about 10,000 medical concepts (and three times as many terms), represented in a dozen concept schemes that reflect their semantic types: anatomy, healthcare professionals, diseases, institutions, etc., and between which we have created about 30,000 relationships.

ALJ: We have created an ontology specific to our needs, which allows for establishing relationships between concepts according to their semantic types. The relationships are specific to the medical domain, for example, <intervention> performed by <professional>; <condition> treated by <professionals>; <condition> targets <population>, and so on. Depending on the needs of our clients, we sometimes create customized ontologies that interact with our taxonomy to address specific needs that do not apply to all clients in a general manner.

In addition to relationships between concepts, we also make sure our taxonomy is interoperable and aligned with other main reference resources, such as UMLS, MeSH, and so on. We thus have a couple of data properties to map our concepts to these resources concepts.

ALJ: Managing ontologies is very simple with Graphologi. In the Ontology section of the tool, we define a main ontology and others that are more specific to certain clients. For each ontology, we create relationships using object properties and restrict their use to classes of concepts in their range and domain. We also create free text fields with Data Properties. Graphologi allows us to write constraints on the content of these fields, which enables us to maintain quality control. 

Compared to taxonomies, ontologies require less day-to-day maintenance. Once the relations and classes are settled, we seldom make modifications.

ALJ: We will soon start working on internationalizing our taxonomy by adding new languages and localizing it to country-specific medical concepts. Additionally, we have an ongoing project to enhance the natural language understanding capabilities of our product, with the aim of creating clinically-personalized digital health experiences in the most seamless way possible.

Thank you Anne-Laure Jousse for these insights into knowledge graphs at Clinia.

To learn more about Graphifi and their knowledge graph tool, Graphologi, please contact info@graphifi.com