Aggregate predictions from individual predictions

[docsteveharris]

This is part of the modelling MWE in that it's the first time we will show the updated aggregate predictions from all the models in one place Including the 'uncertainty'

https://github.com/HYLODE/HySys/blob/dev/hylode/hyui/demand/src/demand/aggregations.py

[docsteveharris]

we need to add the taps to the LoS estimates to calculate total bed demand

[docsteveharris]

suggestion today that we might move this aggregation step into the back end @zmek will review this would then need @harryjmoss to support instantiating as an API that the front end can call

[docsteveharris]

zella working on this still doing some refactoring related to #139 and then this will be an easier follow-on

[docsteveharris]

no updates this week ping @zmek when we're ready: she can pick this up

[zmek]

Hi @docsteveharris

I have created a refactored set of scripts and models which will make it possible for you to retrieve an aggregated prediction for ICU occupancy, with just these parameters

• a date
• the names of the models you want the function to use (ie to pick up from ML flow to make the elective, nonelective and ICU los predictions). The constants representing the names of the models are usually saved in an environment variable I believe (Jon will know). However, I have added logic to parse these constants to decide which location is being requested (gwb, tower or wms) so these will need to be passed in explicitly.

(Jon - we can discuss whether we should add model version; currently this just picks up the most recent version)

How to call the function is shown in this notebook (with all the subsidiary MLFlow pyfunc and sklearn stats functions saved to python scripts)

Still to do

• add a third model for the ICU patients on the unit, when this is available in ML flow from Tom
• make an API for this function (I am hoping @jongillham might be able to do this)

Also (for discussion on Tuesday)

• the training data for this; I have not been able to refresh my training data defining patients on the elective list. I have been using Jen's code to identify elective patients, which doesn't work for me any more, due to a new version of pandas (I think) - very difficult to debug without poring over her code but perhaps Hrishi could pick this up
• whether and how much time we want to spend evaluating these models (I have done none, which is poor modelling practice)
• how we might create an aspiration version of this (ie comparing busy with less-busy ICU); as discussed last week

[docsteveharris]

blocked for now b/c we need a model for icu LoS

[docsteveharris]

Not blocked if we use the bourneville predictions

• get the API's for the 'taps' up and running #125 and #126
• check the cache etc is behaving for all these
• implement the aggregation code on the front end
• the aggregation refers to the aggregate of the ICU LoS work + the taps

Serve this up as a plotly dash function that reads from the taps and aggregate Los above, runs Zella's aggregation function, and returns it into a store on the sitrep page. The display of the aggregation is then a separate piece of work.