Classification of borderline patients

[shraddhapai]

One point that came up in the discussion was to examine the behaviour of the classifier when a patient's similarity rank isn't too different among the classes (rank_A - rank_B --> 0). It's an issue I can put down in the netDx 'to-do' list. I said that while we currently provided a thresholded answer for label assignment, that we were set up for reporting instead a degree of confidence in classification (e.g. one patient may be 0.8 type A while another might by 0.55 type 'A'). We should explore if looking at the distribution of this ranking discrepancy can improve performance by fixing borderline cases.