I have clinical data, where patients are grouped into three groups. However, when I classify group 1 vs 2 (based on their behavioral data), and apply the trained classifier onto group 3, group 3 is almost always predicted just as group 1.

Now I am interested into understanding if it really makes sense to divide the patients into three groups, and if not two groups would be sufficient. I understand that clustering would be one such method, but I am interested if there are not other, possibly classification, approaches, where it has not be decided in advance how many groups (clusters) there are.

The classifier I am using is a linear SVM. The data is based on how subjects are performing on a bimanual task, for instance what force each hand is applying, what the ratio of these forces are, how these forces change in different conditions. I also do have data on age and gender, though performance is more predictive for the conditions. Predicted are three categories of a motor-disease condition.

Any feedback appreciated

  • 1
    $\begingroup$ What kind of data do you have? What do you group your patients on (age, sex, etc.)? What "trained classifier" are you using? What are you predicting? Please be a little bit more specific. $\endgroup$ – Anna SdTC Nov 21 '17 at 23:31
  • $\begingroup$ .............edited $\endgroup$ – user24544 Nov 21 '17 at 23:33
  • $\begingroup$ I am not fully sure to understand exactly what you are doing. How are you splitting the groups? As a general thing, if you apply a clustering algorithm (for instance, a k-means), you can try different values of $k$ and see at what value your reduction in sum of squared errors is small for an extra cluster, but I think it is more important first to understand why and how you want to divide your patients in groups (I understand "groups" are different from "predicted categories"?) $\endgroup$ – Anna SdTC Nov 22 '17 at 1:09
  • $\begingroup$ according to pre-existing diagnostic categories (and gorups and class labels are the same) $\endgroup$ – user24544 Nov 22 '17 at 8:51

Group 3 can be, e.g., a meaningful subtype of 1 or otherwise related.

Consider 1: Apples, 2: Banana, 3: Pears

A classifier trained on 1 vs 2 will usually predict pears as apples. Does that mean pears are not meaningful?

| cite | improve this answer | |

Your Answer

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