I was looking at the functions under R/mispiAppv2.R and had a few ideas for improvement of the back-end:
- Everything (CPT /distribution builders, plotting logic) is stored under the same file, separating them into modular files might make it more maintainable/expandable? (I'm quite new to R so not sure about the feasibility of this)
- There are some assumptions that could benefit from using actual data/priors, like uniform age probability
- Some equations (dfLR computation, build_lr_distributions, build_roc) allow zero-division that could trow NA to subsequent code, adding validation or a small epsilon could prevent errors and/or blown-up plots
I think the second part will have the biggest impact on model performance, and scraping the web for data might be a good approach to set up priors. The third point might be irrelevant for common inputs, but it might be worth safeguarding for consistent performance on edge cases.
I was looking at the functions under R/mispiAppv2.R and had a few ideas for improvement of the back-end:
I think the second part will have the biggest impact on model performance, and scraping the web for data might be a good approach to set up priors. The third point might be irrelevant for common inputs, but it might be worth safeguarding for consistent performance on edge cases.