(1/6) Interesting discussion on REMAPCAP and AI. I first thought ”ick, an unnecessary sidebar to the real innovations”, then “maybe it’s AI in an overly broad sense of AI”….and now that there are pieces to AI, and it fulfills some well, others poorly. Curious what others think. https://twitter.com/statsepi/status/1248277764542783488

(2/6) I tend to think of AI in two pieces. First is the technical implementation (neural nets, etc.). There are big debates comparing these to standard statistical techniques. The REMAP-CAP model is a GLM. Only technical AI to me if you already think most things are AI.

(3/6) The second piece is automatic learning. Most clinical trials use discrete, one time learning (the final analysis). Some AI applications are one time as well, but many focus on automated continual improvement. REMAP CAP is trying to accomplish this as well.

(4/6) Standard RCTs ask focused/narrow questions. We spend a lot of time between trials on what to do next. Historically, we aren’t automated. If we wanted to investigate many antibiotics, antivirals, durations of treatments, etc. we would run several trials over several years.

(5/6) REMAP CAP is attempting to have an automated algorithm investigate the entire space of combinations. It doesn’t need a neural net to do this and doesn’t use one. But it is achieving the automated, continual learning that we expect in many AI applications.

(6/6) I don’t know the right words here. It’s definitely adaptive, but more ambitious than many adaptive trials. Would be nice to have a clean way to emphasize this automated learning without implying there a neural net or other more prototypical AI/ML algorithm at the heart.