How partners can use BMJ clinical intelligence to complement large language models
Knowledge graphs may leverage, work with, or complement large language models both to extract features, define relationships, and make associations.
This approach is massively scalable and it is exciting to consider how knowledge graphs may work with, leverage, or complement large language models to define relationships, extract features, and make associations. These will all result in deeper insights that will benefit both patients and physicians. Knowledge graphs also give us the ability to project the entire patient data set against the graph and to get a better picture of the patient – a more comprehensive picture than what might emerge from using forward and backward chaining through rule bases. When we see the whole patient state reflected against the graph, it can tell us for example that we expected to know a certain fact about the patient but we do not see it or we do not see that a relevant procedure or investigation has been considered – as it is not seen in the data. This is a patient-centered, data-driven approach – projecting the data against the knowledge graph and getting a snapshot of the patient’s current state, and telling the end user who may or may not be familiar with that patient what are reasonable expectations for their care delivery or care journey. Another feature of knowledge graphs is their ability to handle much more complexity than simple rule-based approaches. Rules tend to be relatively condition-specific and so it is difficult to express the complexity of multimorbidity with rule-based systems. So knowledge graphs allow us to see a more comprehensive picture not only of the care domain but of the patient state that is reflected in the care domain. Rules can lead to a lot of redundant alerts or even inappropriate alerts. It is hard to tailor the alert appropriately to a specific user. For example, a cardiologist may wish to have a different set of alerts than a general practitioner. And a pediatric cardiologist may wish to have different alerts than a general cardiologist – so it quickly becomes complicated and difficult to scale. An alternative is to include models of the end users themselves in the knowledge graph and fine – tune decision support that is relevant to them – in essence both patient and provider- centric decision support. Updating can also be a challenge in rules-based systems. Every time the evidence changes, it needs to be analyzed and potentially incorporated into a range of rules. And because rules are specific, every time one part of the evidence changes, the rules need to be reviewed to see if they need to change as well. The same may occur with changes to codes for controlled medical terminologies. However, knowledge graphs can be updated continuously – and we have a team of expert clinical academics and consultants from around the world (the majority are US-based) to survey the literature and guidelines and to inform the content team at BMJ Group when new practice-changing evidence needs to be incorporated into its clinical information and then translated into computable evidence by the clinical informatics team.
11
Why the BMJ knowledge graph
Powered by FlippingBook