A knowledge graph, or, more formally, an ontology, is a comprehensive description of a domain, including the names of concepts and the relationships between them (read more here): it is a representation of our knowledge of a space. As we begin to organize the knowledge in a domain, we rely heavily on taxonomies for naming and defining concepts, and the semantic relationships between concepts create a network of connections which allows us to infer what things mean and how they fit together.
Far from mere stamp collecting, knowledge graphs are designed to solve real-world problems. Knowledge graphs form the basis for knowledge-driven technologies, whether you are looking for a product recommendation on Amazon or are trying to plan your next trip on Expedia. Importantly, this core piece of technology must be designed with the use case in mind, and the application that I am interested in addressing at Syndromic is related to our search for wellness. As defined by the Global Wellness Institute, wellness is the active pursuit of activities, choices, and lifestyles that lead to a holistic state of health. The underlying question that we, as consumers, are trying to answer is "What is right for my well-being?"
There are many ways to answer this question. We could look for scientific evidence in favor of an action (e.g. why we should take a turmeric supplement), we could look for evidence against an action (e.g. why we should not take a vitamin D supplement). More generally, we could consult reference databases that list things that are right and wrong, based on a standard definition.
However, this approach misses the deeply personal nature of the quest for well-being. We are not merely automata, following sets of rules that govern "right" from "wrong." We are complex systems living within a complex system: we are conditioned to consume multivariate inputs, and our minds learn how to recognize and act on patterns distilled from these inputs. The decision we choose to make in a particular moment, whether it is for our health or our home, our children or our chores, is a byproduct of years of experience. We operate not on black-and-white rules, but, rather, on patterns. When the pattern is immediately apparent to us, often as we gain experience over time, we reason quickly about a situation by matching the situation to a known pattern first, and then supporting our decision with logic. In unfamiliar situations, on the other hand, our thinking is slower because we rely more heavily on deductive reasoning, in the absence of a large pattern bank, to guide us (read more on this here or watch this instead). Our deductive reasoning looks like "If this, then that" and "Since A, then B."
We operate not on black-and-white rules, but, rather, on patterns.
Cognitively, we do not operate on a decision in isolation, and our wellness decisions are no different. If my shoulder hurts, for instance, it is insufficient for me to read that I should simply "ask your doctor" (especially since many of us don't visit our doctors, and about a fifth of us in America don't even have a doctor that we could refer to as "ours" since we switch providers so often). Increasingly, popping an Advil for musculoskeletal pain is also becoming an insufficient solution.
What starts feeling like a sufficient answer to my problem is when I start layering on context. For an example where my shoulder hurts when I raise my arm, a more satisfying explanation might be:
Now, that is much more cognitively satisfying than simply popping a painkiller or getting a massage... for me.
The structure of my satisfaction involves two distinct criteria:
1. School of thought, or belief system: the above explanation relied heavily on principles ("beliefs") from biomechanics and kinesiology. This is not the only belief system in town. A neurologist could have relied on their understanding of the neural pathways of pain to prescribe an analgesic, topical or oral. A psychiatrist could have relied on their knowledge of somatization to posit that my underlying stress and anxiety are leading to an exaggerated pain response. An Ayurvedic physician could postulate, based on their knowledge of doshas and the nature of my pain (shooting pain, as opposed to throbbing pain) that I would benefit from a warming massage.
Each of these schools of thought offers solutions to my problem, and each one will have varying degrees of effectiveness. Going back to our question - "what is right for my well-being?" - the definition of well-being is entirely subjective, and we must be allowed to choose which school of thought we ascribe to. The most comprehensive and parsimonious answer to this question would search across competing schools of thought for the "best" answer: this requires us to be unbounded by doctrine or dogma in our search for solutions.
This requires us to be unbounded by doctrine or dogma in our search for solutions
2. Logic, or reasoning: in cognitive science and AI, the pattern of reasoning I employed above is referred to as "chaining." A series of concepts were strung together to form a logic chain, which allows me to reason about the problem. This is a quintessential feature of successful approaches to wellness: a notable wellness or healthcare practitioner has the ability to take a narrowly defined problem (e.g. shoulder pain) and link it to explanations and solutions that touch on meaningful facets of your life (read more here and here about the cognition behind clinical expertise). The critical ingredient for success is in our ability to link, or chain, concepts together.
Unfortunately, we are not all so fortunate to stumble upon such practitioners, and practitioners, in turn, do not have decision support tools to help them rise to this level of proficiency in bridging disparate disciplines and schools of thought. This is a paradox in an era where most human knowledge now sits on a database somewhere, accessible via a Google search bar. I believe that such solutions should be widely available and easily accessible: everyone should be able to find well-reasoned solutions to their wellness problems, personalized to their belief system of choice. To do so, we need an inference engine that can operate at scale, and the backbone of that engine is a wellness knowledge graph.
We need an inference engine that can operate at scale, and the backbone of that engine is a wellness knowledge graph.
My proposal for a wellness knowledge graph is not new, but existing knowledge graphs have been designed with different applications in mind. Let's explore a few publicly available ontologies, which will help to illustrate these differences.
KBpedia has undertaken a herculean effort to curate vast, online stores of knowledge, including Wikipedia. KBpedia's knowledge graph was designed for machine learning, and more specifically, document classification. With this application in mind, KBpedia's knowledge graph has mutually exclusive classes of information: when these classes are used with natural language processing (NLP) to label documents, this results in a clean partition between your positive labels and your negative labels in your training data. While this technology is powerful, it does not help us answer our question on "what is right for my well-being?", for it does not have key semantic classes, like what constitutes a treatment, how and when treatments are used, or schools of thought.
In contrast, Wellzesta published a relatively focused wellness taxonomy, shown here (source here):
The limited depth of this taxonomy reflects Wellzesta's use case of enabling wellness among elderly populations in assisted living facilities. This is evident in the fitness branch, which has 3 child concepts related to "balance" (balance training, yoga, and tai chi) - important in preventing falls in the elderly - while strength training has no child concepts at all. The absence of subconcepts (or "child" concepts) for strength training is appropriate for their use case, but it does not mean that strength training cannot be subdivided further in general. For contrast, let's look at CrossFit, an exercise regimen that emphasizes cross-functional training and Olympic weightlifting. Below, you can see part of the CrossFit knowledge graph, adapted from the CrossFit Journal 2003.
Click on the nodes below to expand the graph:
You can see that CrossFit utilizes several types of strength training in their fitness regimen, which necessitates subdividing the "weightlifting" concept. In addition, CrossFit organizes the exercises by "modality" to enable "programming," i.e. the daily design of a unique combination of exercises and repetitions that they refer to as their "workout of the day," or WOD. We also see that, while "modality" per se is absent from Wellzesta's graph, some of the concepts that comprise "modality" are there. This illustrates one of the primary challenges of designing an ontology: to create a standard vocabulary that allows different applications and different domains to communicate with each other seamlessly.
In addition, for a wellness knowledge graph to answer our question ("what is right for my well-being?"), we need to be able to reason across disciplines and bridge information. We would need to go further than "modality" and add concepts from kinesiology and biomechanics, such as which muscles are eccentrically and concentrically contracted. The muscle groups would then provide the point of linkage to mobility exercises (here's an early example), a set of techniques that have also become popular in the CrossFit community. Using the right anatomic attributes, we could even foray as far afield as yoga or pilates.
Finally, no discussion of ontologies would be complete without mentioning SNOMED (available here). SNOMED-CT, or the Systemized Nomenclature of Medicine, is the most comprehensive clinical terminology in the world. Therein lies the problem: it is strictly clinical. While its semantics are rich - rich enough to power electronic health records - they are focused on terms used within doctor's offices, which leaves many of the domains that we encounter in our search for wellness unturned. Important schools of thought in wellness, like Ayurveda and Traditional Chinese Medicine, are not included, and concepts related to fitness and mobility are not connected to the anatomic concepts required to make inferences.
By comparing broad knowledge graphs (KBpedia), small ones (Wellzesta's ICW), and popular ones (SNOMED-CT), we can see that each graph's structure, breadth, and depth reflect their intended use case. While some of these graphs (e.g. SNOMED-CT) should certainly be a part of a comprehensive wellness knowledge graph, they are not sufficient to answer our pressing question: "what is right for my well-being?"
We need a new kind of knowledge graph. In my next article, I will dive deeper into the design principles I am using to build this knowledge graph at Syndromic.