0
$\begingroup$

Apologies if I mess something up, I've only used Cox PH once and am only starting with deep learning!

In my project, I'm trying to estimate the best interval for screening patients for some complications of diabetes. Currently they are all supposed to visit the clinic every year- but this may not be adequate for all.

I'm planning to calculate the complication risk in 1 year since the last visit for the whole population (let's call it complication risk-CR), which I want to be equal for all patients. This means readjusting the screening interval for those most and least likely to develop a complication so that the time to the next visit is equal to CR.

However, the issue is that patients have the diagnosis recorded not at the time of development, but at the time of screening. Therefore, my data is interval-censored. I understand that I can 1) use Turnbull's estimator 2) follow the steps of a paper I found, where they use "generalised linear regression with complementary log-link function to allow for interval censoring" (I don't know what that actually means).

Two questions:

  1. could you please explain how the two methods differ between each other and from the Cox PH model?
  2. Is there a way that I could use deep learning models in this? Is deep learning even bothered by interval censoring? I don't really care about knowing the impact of individual variables- as long as it can calculate the risk more accurately, I'm happy (although if it could tell me something about the variables that would be great).

Thank you!

Improve this question
$\endgroup$
4
  • $\begingroup$ For "deep learning" AKA neural networks be prepared for using a very complex sample size calculation, and be prepared for the method requiring an enormous number of events in order to be reliable/stable. $\endgroup$
    – Frank Harrell
    Commented Aug 8, 2022 at 12:23
  • $\begingroup$ Hi @FrankHarrell, what do you mean by sample size calculation? and what do you mean by enormous? will more than 100,000 patients do? $\endgroup$
    – Wojty
    Commented Aug 8, 2022 at 12:33
  • $\begingroup$ What matters is the number of observed actual events. $\endgroup$
    – Frank Harrell
    Commented Aug 8, 2022 at 13:10
  • $\begingroup$ How many predictors are you considering for your model? What's the typical time interval between screening visits in your data? How many "events" (detections of complications) do you have? Can the same individual have multiple "events"? Please provide that information by editing the question, as comments are easy to overlook and can be deleted. $\endgroup$
    – EdM
    Commented Aug 8, 2022 at 17:51

0

Reset to default

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.