identifying possible machine learning approaches

Question

I am working on a Biology study and myself and a fellow colleague have looked at various patient medical images and we have computed a clinical score based on the pathology noted on the images. This score is in the range 0-9 (where lower values indicate remission and higher values indicate severe disease).

On top of this we have worked with some image analysis people to compute some automated image metrics. These are typically based on image texture, some measured displacement of certain anatomical landmarks over time and a few others. In total we have 6 of these image derived metrics.

Now, we want to know if there is a mapping between these image metrics and the clinical scores that we have computed. In the beginning, I tried a simple multivariate linear regression and there does not seem to be any link between them. I am now wondering what other methods I could try to see perhaps if there is a non-linear relationship.

My question is that given this kind of data, which machine learning methods do you think would be appropriate to try next? We have about 300 datasets that we have analyzed i.e. 300 patients where we have the clinical scores and the image based metrics that I described before.

Answer 1

Keeping the assumption that human-derived severity score is the "truth" label about the severity of the disease, then your approach is correct, and the only question left is "is there a possibility that a relationship exist even after my multivariate linear regression fails to spot any significant relation?"

I would say that you should in order

1. post the result of the regression so we can judge whether your conclusions are correct, or at least say what metrics you use to conclude there is no relationship
2. mention whether you tried adding interaction terms and or quadratic/cubic/higher order terms
3. if all that fails, you could try a non-parametric regression method such as k-nearest neighbor (KNN).