Support for on-demand predictions

[davidaastrom]

Thank you for this package! I have been exploring how we could use it at Karolinska, Sweden and found some features that would make it more useful for us.

Is Your Feature Request Related to a Problem? Please Describe

When exploring how we could use this solution we found that the user interface options that we would want are not currently supported by the package. The information sources our staff uses for this type of decision are web-applications and BI-tools that are based on near-real time data. To have the possiblility to present our predictions in the same interfaces, they would need to be based on the same data and not limited to a set of pre-determined prediction times.

Describe the Solution You'd Like

I would like to generate predictions for demand in a time window at an arbitrary time. I believe this would mean to support replacing the ed-admissions models with one and providing time/hour of day as input to the models instead.

Describe Alternatives You've Considered

No response

Additional Context

No response

[zmek]

@davidaastrom Thanks for this excellent suggestion. UCLH has the same aspiration for on-demand modelling.

In terms of the models themselves, I think it's fairly straightforward.

• Currently we train a classifier for each prediction to predict admission or discharge. But it would perfectly possible to use a single classifier to predict admission after ED, and introduce a time of day feature(s) to allow for variability over the day
• The models predicting specialty of admission, as currently implemented, are time-of-day agnostic
• The models predicting yet-to-arrive distributions are time-of-day agnostic since version 1.3.0

The trickier part will be working through the rest of the architecture so that you can run the full prediction pipeline at an arbitrary clock time against live patient state; this means on-demand snapshot/feature construction, loosening the prediction-time-keyed model bundle and evaluation APIs.

We share this goal and hope to work on this in the future.