The FHIR standard started as a better way to exchange healthcare data, but it also provides a solid basis for deep analytics and Machine Learning at scale. This post looks at an example from the recent FHIR DevDays conference that does just that. You can also run the interactive FHIR data engineering tutorial used in the conference yourself. Our first step is to bring FHIR data into a data lake – a computational environment where our analysis can easily and efficiently work through petabytes of data.

We’re excited to open source Bunsen, a library to make analyzing FHIR data with Apache Spark simple and scalable. Bunsen encodes FHIR resources directly into Apache Spark’s native data structures. This lets users leverage well-defined FHIR data models directly within Spark SQL. Here’s a simple query against a table of FHIR observations that produces a table of heart rate values: spark.sql(""" select subject.reference person_id, effectiveDateTime date_time, valueQuantity.value value from observations where in_valueset(code, 'heart_rate') """).

Over the past few years, Cerner has been developing a standards-based platform on top of our Cerner Millennium® EHR. Rather than roll our own API, we’ve been using the HL7® FHIR® standard. For app integration, we’ve been using the SMART® on FHIR specification. If you’re not immediately familiar with those acronyms, you’re definitely not alone. As we were developing the services, we were also fielding questions from 3rd party developers, answering a lot of questions internally, and trying to keep up with the specifications themselves.