I am developing a ML model where i am working with medical specialties.
My dataset contains different specialties (General surgery, urology,etc...) . First I build a model with all my data with the specialties mixed, and then i apply the same ML algorithm at a specialty level. I achieve better evaluation metrics for all the Specialty models except for one specialty can someone help understand the reason?
What do you mean with "Performs worse"?
Because in ML there are 2 phenomena:
- Overfitting: When you predict so, but so well the training data, that when you input new data (Test data) it won't predict because it can only predict the training set.
- Underfitting: When you train your data so poorly that there is not enough information to predict a test set.
I believe your issue might fall into the "Overfitting" phenomena but without any further knowledge of the problem you want to solve, it is hard to say.
Hope this helps.
Regards,
Juan
Dear Alice,
as far as I know, there is mostly more than one image from each patient in medical datasets. Regarding that, while splitting your dataset or working with a small subset of the whole dataset, you should pay attention to samples to be such that:
- No two images (scans) of the same patient are presented in the train and test set simultaneously
- [Softly] keep the ratio of classes
So in your case, there is a chance that the subset you've chosen and trained your model on, doesn't contain enough number of samples of that specific specialty.
I hope it helps you!
That one speciality might be an outlier in terms if it’s properties. So when you apply your general model to the specific outlier, it performs worse.
- Badges
- Science topic
- Similar topics
- Psychology