I'm thinking of an application for diabetics, that, given previous values of blood glucose and insulin dosage, predicts the glucose level for the next few hours.

I know a few things about neural networks and perceptrons, but not much. And there are probably whole other worlds of other machine learning methods. So I'd like to ask about what way should I try to go.

My problem is below:

The app would probably have a (very) simplified model over what is the reality (what food, how much of it, did some sport lately?, what kind of insulin, etc.) - and there are also some "expert knowledge" rules that I know but that wouldn't get programmed into it (eg. if you have too low glucose level, your body compensates and makes it into a too high glucose level). I want to try how far can I get just with the glucose measurements (when and how much) and insulin dosage (when and how much).

I guess the main question it would be nice if the program solved is "Given my glucose history, what glucose will I have in a few hours if I take X units of insulin?" (That could indirectly solve question "How many units of insulin should I take if I want to have a good glucose level in a few hours?" but that's more complicated matter.)

The food is an important part of the question, but I think it's closely linked to the insulin - the two balance out: if you eat food, you take insulin. You don't take insulin without eating (it's more complicated than that, but I'm trying to make it simple), so I guess it could work without putting in the data about food.

Now the data (at least timewise) isn't uniformly distributed. Ideally it should be (regular measurements at the morning, noon, evening, etc.) but more often than not the measurements get skipped. So the training dataset would have to deal with having "holes" in it. I guess that's ruling out some methods.

What do you recommend to try? (Again, this is not anything medically professional and "to be used in real life" - it's just a proof of concept for me and a toy project to try to do some machine learning.)

  • 4
    $\begingroup$ Don't get me wrong, but I think you need to understand a little more about the regression model before you dive into the application phase. You will need to choose your features and your model. A good place to start will be Linear Regression and then pick up logistic regression and SVM. $\endgroup$ – Subhayan Sep 7 '13 at 23:41
  • $\begingroup$ @Subhayan Yes, I think that's also what my question is about (what model, etc.). I will look into your links. Thanks! $\endgroup$ – Martin Janiczek Sep 8 '13 at 8:13
  • 4
    $\begingroup$ If you don't have any prior experience in Machine Learning, may I suggest taking a look into the first couple of chapters of The Elements of Statistical Learning. And if you have worked on ML before, a very powerful model to treat sparse data would be Relevance Vector Machine (see this explanation too). Also you can take a look into Online Algorithms if you want to make a more professional application. $\endgroup$ – Subhayan Sep 8 '13 at 9:37
  • 4
    $\begingroup$ Another ML text (with an information theory focus), available free: inference.phy.cam.ac.uk/mackay/itila $\endgroup$ – András Salamon Sep 8 '13 at 15:15
  • $\begingroup$ Do you have a data set from a glucometer to start? If not, It might also not be a bad idea to get a sample data set and begin analysing it so that you know what you are working with. $\endgroup$ – Mr. Concolato Jan 7 '15 at 22:28

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.