There was nothing special about the tonsillectomy cohort. The fact they were were unexacting was exactly the reason we picked them for the feasibility study. We carried some rather plain priors into this study, primarily that there was little variation in anaesthetic practice and that it was of little consequence.

We were pretty close but the data infrastructure we built along the way was the really exciting bit. There was nothing special about the tonsillectomy cohort. The fact they were unexacting was exactly the reason we picked them for the feasibility study. We carried some rather plain priors into this study, primarily that there was little variation in anaesthetic practice and that it was of little consequence.

We were pretty close, but the data infrastructure we built along the way was the really exciting bit.

Aims and Background

As part of the OPALS project we declared that we would build a highly granular peri-operative dataset that would allow us to now test questions relying on both detail and breadth. As part of building this pipeline we needed a feasibility study cohort to test the data extraction and methods.

As part of the OPALS project, we declared that we would build a highly granular peri-operative dataset that would allow us to test questions relying on both detail and breadth. As part of building this pipeline, we needed a feasibility-study cohort to test the data extraction and methods.

Tonsillectomies were used as the cohort, as their intra-operative analgesia relies exclusively on medications that are tracked within the anaesthetic chart data tables (IV and oral meds). There were no spinals, blocks, or PCAs to worry about. We were also interested in how the coblation technique would interact with the analgesia and pain.

From a feasibility point of view, they were also a nice, easy cohort to define and fairly common, with our final analysis having over 3,000 encounters.

Work

Data extraction

This project really got to grips with the EPIC data model within the UCLH EHR. I spent a lot of time building (and rebuilding) SQL queries to get exactly the data I wanted, and the data I was getting was what I thought it was. Whilst the data model developed was later eclipsed by IMPACT, the lessons learned were key in understanding how peri-operative interactions translate to data.

Analysis

We had initially planned an IV analysis whereby we would investigate the level of variation by clinician, then stratify and investigate the proximal outcomes. In the end we decided to limit our analysis to a descriptive analysis owing to the lack of any real variation in practice. It did give us the opportunity to think about what could make a good instrument in a similar cohort, though.

One really cool bit was the UpSet graphs we got to produce. We came across the problem of displaying proportions of combinations. When there are three binary decisions we can produce a pretty good-looking Venn diagram. However, once you start to get to more than five combinations, it becomes very tricky to read.

I stumbled upon the UpSet website through some frustrated googling, and the results came out looking pretty good!

Results

The ultimate output was a paper (currently in draft), which will hopefully be my first full-length, first-author paper.

We then took the framework from the code developed and used it in our future projects on the colectomy cohort.