Automating the clinical coding process in an NHS Trust — ep. 3
Thought turns into action
A while ago I began this short series by talking around the broader concepts and considerations relevant to automating clinical coding at Kettering General Hospital (KGH), part of the University Hospitals of Northamptonshire (UHN) Group, before a period of virtual ‘radio silence’. Suffice it to say this period of relative quiet belied the actuality of feverish work…indeed, as I write this we find ourselves at the end a fruitful project with colleagues from TPXimpact and we’re live with a functional automated solution busy codifying all endoscopy daycase patients at KGH. I feel like now’s as good a time as any for a short informal retrospective and a quick look ahead.
A brief recap
In brief from a clinical coding point of view daycase endoscopy is a high volume, low complexity venture. Coders work to monthly deadlines for data submissions and “endos” are the numbers game; work to be coded daily in between their more complicated or longer stay patients. They’re not particularly challenging or engaging for an experienced coder but do take up a lot of coder time in totality. Endos then were a first good potential area to dip our toe into automation.
In addition to the above endos involve a sufficiently complex patient case mix with a reasonably limited procedural scope. The depth/number of diagnostic codes typically assigned will vary between 1–6 codes depending on the patients comorbid status and their relevant personal/family medical history. And yet a total of only 6–8 procedure types form the vast majority of interventions within the service.
An area then that were we to automate would have an immediate benefit for coders by freeing them up to code more complex patients and a challenging enough diagnostic picture with a reasonably limited number of interventions. Not only that but owing to the diagnostic depth any learning throughout this build process could be turned to future projects thus reducing their lead time.
Enter “IanBot”
We named our automated solution IanBot as a leaving thank you to our colleague Ian Roddis.
IanBot is primarily a robotic process automation solution as we felt this was most appropriate for our initial ambition and the coding casemix we were looking to automate. However we also established a process for iterative learning within the solution, where it could learn from previous examples. This reduced the number of exceptions — instances that the system cannot automatically process — and improved the speed and accuracy of coding. This iterative approach saw the creation of a ‘broader’ automated solution that impressed us with it’s output.
Obviously a lot of work went into standing up the concept and managing the overall project but when it came down to really building and fleshing out what was to later become IanBot two individuals were crucial to this — my colleague Liz Ivatt working as the clinical coding subject matter expert (SME) with Grant Grobler of TPXimpact developing Ian. The two worked via an iterative approach during their build phase each also learning a bit about the other’s respective fields as they went — this working dynamic Liz feels was critical to the success:
As the build began it was obvious to the developer and the SME that we could up the bar and teach the AI more complex procedures. Not only were we teaching the AI, but we were also guiding the developer; Grant needed to understand the Clinical Coding standards so he could instruct the bot how to assign the correct primary, and secondary diagnosis and comorbidities, as well as learning the standards associated with Endoscopy procedures. This confirmed to us that it is crucial when building a Coding automation system that a Coder work closely with the developer.
Tweaks and adjustments continue however as of writing IanBot currently automates on average 87% of our monthly endoscopy activity with an average accuracy of primary diagnosis & primary procedure assignment of 94%. In addition to the primary diagnostic/procedure codes Ian also assigns appropriate codes reflecting patient’s documented comorbid status and medical history.
Looking ahead
Our experience working on IanBot has clarified some of my thoughts around automation in the clinical coding space within the NHS. I can see now a strong argument for the creation of a clearly defined post within a coding service that focuses on the development of automation across the function, acting as a subject matter expert to liaise with programmers and engaging key stakeholders across the healthcare picture.
As we further automate future processes a coders role will grow and evolve — we’ll be gently introducing these skilled professionals to new opportunities that they typically wouldn’t have access to until later in their careers — auditing and validating proportions of the bot output against national coding standards and potentially broadening the scope further to include a working skillset with a programming language such as Python.
With our work with IanBot and our plans for the year ahead we’re placing our first banner down, our first signpost for ourselves and others to collaborate with — ‘this way to a Digital Clinical Coding service’.
With thanks to Liz Ivatt, Grant Grobler, Ian Roddis and Andy Callow.
