Perioperative Prediction Suite

[Evan_Minty]

Hi -

I’ve alluded to this project in other posts, but thought I would make it a more official call for network collaborators. This study would be coordinated out of Stanford, I’ll be working on it with a colleague in the BMI program (Lichy Han), and Nigam Shah.

Perioperative medicine is an area where many prediction models are used in practice. Like others PMs in medicine, parsimony is generally the rule here, they can likely be outperformed statistical learning methods operating on a comprehensive representation of the health record like the CDM.

These are largely used to inform risk discussions with patients, but also to direct post operative monitoring and testing (in new Canadian guidelines, decisions around post op Troponin testing can depend in part on the 6 point Lee Cardiac Risk Index for example). They could also have a great use case in directing resources (not all folks need to see an internist pre or post operatively).

It’s an attractive area for the OHDSI PLP framework: these are used clinically in current state, the outcomes are relatively proximate, and we don’t need to try to handle granular data from the hospital stay when the event occurs. In fact, these are more typically used in a preceding outpatient preoperative evaluation.

There’s a number of outcomes of interest which we will look to specify. The most impactful would be MACE/Death.

My current problem, is actually trying to wrap my head around some preliminary results that are too good (AUCs 90% or higher). I maintain a suspicion that this is the result of features from the inpatient stay where the outcome event occurs, making it into the model. Hoping to resolve that issue in the cohort definitions with this thread:

PLP: Restricting features that co-occur in the same encounter as the outcome

Hi, Many outcomes of interest in prediction might occur in the context of an inpatient encounter. Within the current ATLAS framework, is there a way to only use features that occur up to (but not including) the visit where the outcome of interest occurs? If not, I think this would be a useful configuration to add to the framework. The issue is that (probably due to erroneous / insufficiently detailed timestamps), other procedures, diagnoses, etc. that co-occur in the same inpatient encounter…

(In full disclosure: Lichy and I are in a class where the output is an R package - what better than to produce a network study! We’ll make clear academic distinctions about our contributions vs. those of collaborators, our intent to do this has been proposed and approved by the prof teaching the class). There will still doubtlessly be work involved in packaging this up.

[Mark_Danese]

Just to ask an obvious question, but can you edit the analysis data set you are working on? Can you get dates for all of the codes and drop those with the same date as the index date?

[Evan_Minty]

Hi Mark -
Thanks - I was wondering if it would have to ‘go there’ myself. Some of our helpful community members have pointed to some time window offsets in the featureExtraction package (in the thread I alluded to). I’m pretty sure this can be configured within the current framework. I have to spend some more time with the documentation.