When designing the dose recommendation API from scratch, I always imagined distinct steps:
Specify the search space: what parameters can we vary? Some are continuous, others are discrete. This defines an n-dimensional space.
Define evaluation criteria for each point in n-dimensional space. This may be 'CONC - TARGET', or may be 'Clinical Utility', 'Cost', etc.
Define 'selection' criteria. Which treatment regimen will you eventually pick or recommend. This is often either "f(X) = Y" or "max f(X)", but may combine several "Y > Z, and max f(X)".
We should evaluate the API we have built so far, and check if it fits the paradigm above.
When designing the dose recommendation API from scratch, I always imagined distinct steps:
We should evaluate the API we have built so far, and check if it fits the paradigm above.