How to select algorithm model for PRS score predict disease?

Question

[li]

I have recently been using PRS scores for disease prediction, after calculating the PRS score using Ldpred, I found a problem that bothered me, that is, why almost all the literatures used logistic regression to construct the PRS prediction model with age and sex as covariates, instead of using machine learning methods such as SVM, random forest, decision tree and KNN etc? Who can give me some explanation? Thank you very much!

Answer 1

[Kevin Blighe]

One good reason is that 'machine learning' methods are horrific in terms of their applicability to new data. They are designed to work exceptionally on the dataset on which they are trained, and there will likely be a testing / validation dataset that is of the same nature as the one on which the algorithm is trained (thus, they are not actually great for validation).

I write more about this here: What is the best way to combine machine learning algorithms for feature selection such as Variable importance in Random Forest with differential expression analysis?

Let's not get carried away with these buzz terms.

Kevin

[li]

Thank you very much for your explanation! I have read carefully your another answer mentioned above. So now I only need to do is to find the factors that are most relevant to this disease and use logistic regression such as 'glm' to build a prediction model. Do I understand it correctly?

[Kevin Blighe]

That is one way [of many ways] to do it - yes. Nobody can really say, when commencing, which method will prove 'best'.

[li]

Thank you for your reply. I will start to try different methods.