In one work, suppose multi-disease (skin cancer, covid-19, Monkeypox, and lung disease) data analysis and disease prediction, I used 5 ML models (KNN, GB, RF, SVM, and Ada boost) after pre-processing steps (data cleaning, encoding, missing value handling, outliers handling, feature engineering, etc.) now while analyzing the model performances I want to use statistical test on our ML models. I found no previous works that work on the same kind of dataset instead most of them worked on similar types of disease (i.e. Lung cancer (3 types) = 11, 12, 13) detection. I want to know, (1) How can I compare my work outcomes with others as they not using similar data? (2) When applying the statistical test will be appropriate? (3) Why should I use it? (4) How can I perform statistical tests for my work?
it sounds like you need something like a middleware that converts some of the factors you are analyzing to have some sort of a commonality index. So at least there are some basis on why you compare them that way. Specially if like you mentioned there is no existing studies published yet about this approach.
Do some factor analysis to reduce the number of things being measureed to some similar ones for comparison purposes.
Also ensure your instruments reliability suffices' chronbach's alpha and employ divergent and convergent validity tests so that at the end of the day, you are:
1. Measuring the factors that are supposed to be measured
2. Excluding the factors that are not supposed to be measured.
I think applying these principles to your tests could help you create a new approach. Since its a new territory to be explored.
Just my thoughts on this matter. Good luck!