Just “attended” this conference - last week. While some sessions were very techinical to R users, a number of the sessions were high level how systems have implemented algorithms both home-grown and commercial algorithms in a risk-mitigated process. I still have a few more talks to get through - but they state registration still open to gain access to recorded talks and that some talks will be freely available in the near future (US$50 registration fee).
The talk Bringing Machine Learning Models to the Bedside by Karandeep Singh was excellent - worth anyone thinking in this space to see the data-curation, model development, testing, testing, testing, implementation, and governance that University of Michigan has in place.
I think part of the left me gobsmacked information in the talk - to implement proprietary algorithms (Epic) the cost of the annual subscription is US$400,000. The information provided in advance is essentially a single sheet - model information sheet - which he presents another paper showing that 50% of the TRIPOD items are missing. So, part of the 1st if not the whole 1st year is spent silent running, calibrating to local environment, and just seeing if annual subscription fee is justified by model performance in the health system you operate in. Excellent examples from the Epic equivalent of EWS showing that the performance was highly variable, required a lot of training on apppropriate use of the information, and ongoing monitoring and evaluation for calibration drift. Also highlighted that when new events (black swans) arise, the models trained on old data essentially crash.