What is the best activation to use for a keras NN predicting risk of a single binary outcome? Is it sigmoid? And are there some approaches I can use to get better predictions on the low-frequency risks?


I have about 400K records of surgical procedures and multiple post-procedure adverse outcomes. I will train separate models for each outcome.

The risk of these outcomes varies from ~1:10000 to ~2:10. (I recognise I may need to group some of the rare outcomes together)

I'm hoping to use a keras to produce a predicted risk based on various factors (such as patient age, medical conditions, class of surgery, etc etc). I can then use the predicted and actual rates in various populations (eg patients treated by particular doctor) to look for patterns of variation.

I have this example here, as well as this paper (model) both of which use the sigmoid activation.

(As an aside, the situation is similar to the ACS NSQIP Surgical Risk Calculator but the my goals and data vary from this tool.)


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.