I think what we’re arriving on is two things:
- Issues in dealing with temporality.
a)There may be some dimensions of data (concepts derived from billing codes or procedure codes for example) that are applied to an entire encounter and would need to be excluded. I’m actually not sure how the dates are handled. In working through an ETL locally, we’re creating ghost encounters for some billing data as we don’t necessarily know what encounter they belong to if the record is complex.
bottom line is, if a procedure code for intubation or a diagnosis code of respiratory failure can become a predictor in your model, this won’t be doing anything interesting.
b) Consequently, you’d probably only want data sets with temporal and clinical features on the order of the electronic health record. Don’t know what that limits you to in the network.
- (1b) would probably be needed to outperform tools like CURB65, most of the criteria there parameterize those ‘gestalt’ features (confusion, end organ dysfunction / inflammatory response etc ) that probably predict an ICU admission.
You could try to predict ICU admission based solely on previous features in the health record, but you’d lose information that correlates to the severity of their current presentation. That will be a tough ask, but you could give it a shot…