Last year I was privileged to meet with informatics teams at seven influential EMR software vendors. Part of the agenda was to explore the challenges they face when using SNOMED CT. You know the refrain: SNOMED is “too big, too complex, too difficult, too costly” and there is “no incentive to implement it”. As part of a renewed focus on overcoming “barriers to adoption”, I wanted to gauge whether these issues persisted, or whether the situation had changed.
As it happens, things have changed. The focus is no longer on SNOMED’s complexity as an obstacle. It’s on how the complexity inherent in SNOMED’s semantics can be harnessed to deliver problem solving insights from years’ worth of stored clinical records. The perceived clinical and business benefits from the analysis of clinical information are powerful incentives to leverage SNOMED CT’s inherent structured clinical knowledge. Major EMR software vendors now have ontology teams focused on developing machine learning algorithms to structure and mine clinical data, leveraging large corpi of clinical data that have amassed over the past decade or so. Where SNOMED was not used to structure clinical information at source, natural language “tagging engines” can now apply SNOMED structure retrospectively. SNOMED CT is very much part of the momentum gathering around clinical analytics.
The good news is that we can move on from SNOMED being “too big, too complex, too difficult, too costly”. Not a single vendor corroborated this view. But some earlier uncertainties persist. How, for example, to know you have the right SNOMED content to represent a given domain, and that it’s accurate and complete. And it’s sobering to realize that broader adoption scales up the problem space. Previously, EMR vendors were challenged to integrate and manage clinical terminology and classifications in their applications and at their customer sites. Lately, designers of analytics algorithms must demonstrate that their results are repeatable and verifiable given the same kinds of information from multiple disparate sources. It is right and fitting for software to rely on SNOMED for structure, and machinable clinical knowledge. That’s why it’s there. But success depends on consistent, standardized semantics within SNOMED, and in the way SNOMED is used in software across systems, organizations and national boundaries.
Added to this, Informatics teams tend to be under-resourced. “We are constantly reminded that we are an EMR company, not a terminology company,” one informatician recently remarked. How do informatics teams, whether at EMR vendors, national terminology centers or elsewhere, achieve some level of confidence that they have the right content, apply the semantics to clinical information in the right way, and manage the terminology and classification environment efficiently and cost effectively? In the next few weeks I will share some insights gained while visiting those vendors, along with some thoughts on scale-related challenges that face SNOMED users, and how they might be overcome. As always I am happy to have feedback so that we can benefit from one another’s experiences.