Te Whatu Ora exploring AI for clinical coding

eHealth News
,
1

NEWS - eHealthNews.nz editor Rebecca McBeth

This is a companion discussion topic for the original eHealth News article:

https://www.hinz.org.nz/news/663147/

2

This might be a little confronting for clinical coders; “I am going to be replaced by a robot?”. That said, it’s an opportunity to for professional growth, and we will need experts to remain in touch with the complexity, analysis and interpretation of data. If you don’t already know about clinical coding, I highly recommend getting to understand the basics. It’s just as complex as the healthcare itself!

Will AI be able to ignore the acronym of AS, typically Aortic Stenosis, and know that in the context it was made, the clinician was meaning ‘antegrade-stent’?

Clinical coders & Health Information specialists, have a wealth of knowledge and hopefully we use this to scope and integrate AI into clinical coding workflows. Historically, we haven’t been interested in improving the quality of clinical coding, unless it ticks the finance box, so I am hopeful we are seeing the bigger picture. There is so much unlocked potential in coded data. Many countries code outpatient records, but we don’t. Perhaps we might start there.

Is AI-assisted coding going to help, if the medical record or note is of poor quality to start with? :thinking:

4 Likes
5

We’ve been automating repetitive processes using IA/RPA across the metro DHBs for a while now. Our general experience:

  1. There’s been no evidence of FTE replacement (or release).
  2. Robots augment humans, i.e., humans now work on the exception cases (5% to 15%) that the robot is not trained to automate, or there is diminishing cost/benefit if it were to be trained.
  3. Overall, there is an uplift in quality, turnaround times, and processed workload volume without increasing the human resource footprint.

#2 ensures there is limited loss of institutional knowledge.

My MBA engagement was working with the EHT team on the automated mortality coding project, i.e., discovery work to see if NLP was effective in identifying and correctly classifying the underlying cause of death.

Findings:

  1. The quality of the death record was a limitation.
  2. Quality couldn’t be managed without all the things you mentioned in your post, i.e., humans would need to monitor exception cases and also on-going model and data drift.
  3. Overall, the NLP model correctly identified and classified up to 75% of the deaths from the sample.
3 Likes
6

Thank you Parag for replying and giving some more context to the headline. My empathy was for any of our kaimahi, whom read this headline and think their jobs are imminently threatened. So true the idea of robots augmenting humans. Thank you for sharing your MBA thesis, very interesting. And lastly, that quality part is key, yet (broken record) we spend very little effort on improving documentation quality in clinical areas. Best wishes.

1 Like