We need to add some concerns about the data we’re getting and the methods by which we get that data.
Scenario: Receive HL7v2 over MLLP
You might start out with some kind of pipeline that looks like this. You read the data in, parse the HL7, translate fields to something your database understands (M -> Male, F -> Female, for example) and then store it.
Parse the HL7
The best open-source option for parsing is probably HAPI, and even that has you doing most of the work when it comes to things like character encoding.
Even if you nail your parsing, you still have to deal with changing data from other systems. An upgrade might switch where data is sent, or new segments will be added without you realizing it.
Let’s say that you’re successful with this, and your team is on board with the HL7v2 model and can adapt to change. Next, you need to tackle the consistency problem.
Normalize the data
Take a look at the PID segment specification from chapter 3 of the HL7v2 specs.
If you were to guess where to find the patient identifiers, where would you look?
If you guessed either PID-2, PID-3, or PID-4 you wouldn’t be wrong. At one point in time all three were valid. Now, PID-3 is recommended for all identifiers. That doesn’t stop people from using multiple fields, with repetitions.
You can see where we can solve this consistency problem on a case-by-case basis. But how do we solve it more broadly, in a configurable way that doesn’t slow down your implementations? Redox has its own DSL that we just call “Configurations” that solves this for us, but it still involves a lot of communication when we’re bringing new health systems online.
Fill your database
Once you’ve established a pattern and tools for making the data coming into your system consistent – you still need to file it away.
How do you represent clinical concepts internally? Do you use HL7’s models for how things like orders and observations fit together? What kind of process does your team follow when adding new fields to say, the Patient table? Who decides what set of codes should be used to store allergies, medications, and problems?
HL7 makes recommendations for some of these things – for example see Medication in FHIR. But is that list of SNOMED codes going to be the best for your end users to work with? How do you represent the need to do one-to-many types of mapping between codesets?
How are you going to manage when codesets change, like the ICD-9 to ICD-10 transition. Are you going to have to migrate massive amounts of data to the new format?
The Redox Plurality of Concerns
From the very beginning we’ve had a vision of a healthcare network where many of these concerns were tackled by a central hub. It’s been very exciting to build tools around solving these problems, and I’ll be the first to admit we’re only scratching the surface.
I’m just going to end this with Dharma’s conclusion:
This is the work we do at Redox. The good news is that we are thinking about the plurality of our concerns, and we do that as upfront in the design as possible. Meeting all of these expectations is both daunting and thrilling. It’s the nature of our space, and it’s why we do what we do.