Welcome! Please contribute your ideas for what challenges we might aspire to solve, changes in our community that can improve machine learning impact, and examples of machine learning projects that have had tangible impact.
Lacking context for this site? Read the original paper: Machine Learning that Matters (ICML 2012). You can also review the slides.
Finding new categories of patients to improve treatment
  • In many medical domains, physicians do not understand what leads some patients to fare worse than others.  I would like to see a machine learning applied to address this issue.  The measure of success, is if the results are then followed by medical research that figures out how to help these patients fare better.
  • 3 Comments sorted by
  • I agree.  There are a number of diseases that almost certainly have real sub-categories/sub-types that are not well identifiable from symptoms alone.  (In the work that I've done with neuroscientists, schizophrenia has this flavor.  There are very probably a number of sub-types of schizophrenia, and schizophrenia even as we currently recognize it appears on a spectrum that passes through bipolar and into depression.)  I think that a lot of physicians feel that they could do better treatment if they could identify those sub-types, diagnose them, and condition treatment on them.

    Another direction along this line is genetic-based personalized medicine.  That's getting a lot of action in the medical community, but it's a *huge* data analysis/ML problem.  Very high dimension, very low amounts of data.

    And, of course, there's also a huge number of hidden variable issues involved in the "why do some people fare better/worse under treatment X".  Everything from that person's genetic history, to their diet, to the particular cohort of intestinal bacteria they have.  Big ML problems here, but very, very hard to ground truth.
  • One small step in this direction was just presented at ICML 2012:

    A Hierarchical Dirichlet Process Model with Multiple Levels of Clustering for Human EEG Seizure Modeling

    Drausin Wulsin, Shane Jensen, Brian Litt


    I spoke with the author at the poster session and asked if he thought his system could, as a "stretch goal", possibly identify a new group of patients, one that perhaps the physicians had not noticed before.  He hoped that would ultimately be the case (with more work).

  • I feel uplift modeling can be useful here.  It allows for taking a control (placebo) group into account and finding cases where the effect is high with respect to the control group.  This is what doctors usually want.

    See e.g. my paper on ICML'2012 ML for clinical data workshop


    or this paper


    I would be happy to apply this to somebody's data and get real feedback.


To post or add a comment, please sign in or register.

Tip: click the star icon to bookmark (follow) a discussion. You will receive email notifications of subsequent activity.
If search doesn't work, try putting a + in front of your search term.