Many machine learning studies do not have separate test/validation sets and only report cross-validation results which as far as I know are averages of multiple tests (depending on the number of folds). Therefore, true positive, false negative, etc. data cannot be calculated from the reported sensitivities, specificities, etc. (For example, in a study of 100 cases with 2-fold cross validation, if one validation instance results in a 49% sensitivity and one in 50% sensitivity, the average sensitivity would be 49.5%, meaning TP would be 49.5 which is not logical). Should such studies be excluded from meta-analysis? If not, how should the data be extracted/calculated?
Any Physical world that can be modelled as Mathematical world in the form of input to output representation is nothing but machine learning model that can be defined on some data at hand. For meta analysis, you may have look at statistical models Article Statistical models for meta-analysis: A brief tutorial
Statistical regression models can be use for meta analysis as well. Some related work to your query is : Article Functional Connectivity of the Corticobasal Ganglia–Thalamoc...
Article Meta-analyses of human gut microbes associated with obesity and IBD
Dear Ali Nowroozi
Follow these steps:
1. Set aside a portion of the sample data-set.
2. Train the model with the rest data set.
3. Use the reserve portion of the data set to test the model.
For more: https://www.geeksforgeeks.org/cross-validation-machine-learning/
Kind Regards
Qamar Ul Islam
Dear Qamar Ul Islam
I was actually asking about including studies in meta-analyses.
OK so based on the available literature, I found out that there are two approaches:
1. Include the averages and calculate TP, FP, etc. based on those.
2. Only include studies without cross-validation in quantitative analysis.
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Meta-Analysis