6. Enabling rapid model deployment in healthcare, Felix Brann, Vital

https://www.youtube.com/watch?v=HWD42AHrgZk&ab_channel=Tecton (opens in a new tab)

• Motivating exemple: Wait time for patients.
  • No info except grabbing a nurse or a doctor
  • ML to help inform patients during their stay

• Cold start problem: scarce data

  • Hospital expect accurate ML results from day one

  • No history when opening a new facility

  • Emergency departments varies a lot

  • Data regime change (think Covid)

    

• Solution: a facility-agnostic model, predicting the wait percentile

  • Model use raw feature and aggregated features at the facility level

  • Instead of predicting in minutes, we predict in percentile of the historical wait CDF, of a given facility (see below that it can varies a lot for same percentile)

    

• Implementing using Tecton

  • Feature service for all feature to normalize, indexed using a facility_id
  • Lambda periodically extract tecton features into Redshift
  • Redshift is the data source