Discussion: View Thread

  • 1.  Literature References Sought

    Posted 08-23-2021 17:28
    At the risk of asking something a bit off the wall and/or potentially not very relevant to the section:

    I am in search of any paper/study that has applied a scoring/classification approach commonly used in traditional modeling for credit risk, customer churn, etc., to disease progression, in which the target bad/good is defined by presence/absence of progression over a fixed period of time rather than by time to event. If anyone knows of any such thing, I would really appreciate the reference. I have not been super successful, in part because I tend to work more on business than on research.

    Bonus points for oncology :-) Please feel free to email me at michiko.wolcott@msightanalytics.com. Thanks in advance.

    [EDIT] To clarify: I'm not looking for the details of the credit scoring or customer churn modeling and techniques per se, but rather examples of applying a similar framework/approach in predicting disease progression, essentially treating disease progression as a predictive (binary) classification problem. Apologies if my original post was not clear.

    ------------------------------
    Michiko Wolcott
    Principal Consultant
    Msight Analytics
    ------------------------------


  • 2.  RE: Literature References Sought

    Posted 08-24-2021 00:14
    I take it you don't have a binary event as an endpoint.

    Are there multiple occurrences of the same event?
    Are there quantitative amounts that accumulate constantly?
    Are there random amounts that occur at regular times and accumulate?
    Are there fixed amounts or random amounts that occur at random times and accumulate?

    How do your observations come to an end? Do subjects drop out or do they go on to the end of your study period?

    ------------------------------
    David Smith
    Statistician
    ------------------------------



  • 3.  RE: Literature References Sought

    Posted 08-24-2021 00:48

    Yes to all, which are all being dealt with separately and in parallel. What I'm describing is just one piece. Given a fixed observation period (say, 24 months), a patient is considered "bad" if the disease progressed at least once, good if there is no progression. Then it essentially becomes a binary predictive classification problem (discriminant/logistic/other supervised classification; more interested in the application of the framework than the specific method). This is actually a quite common approach in other fields (like credit scoring where I've spend a lot of time) and from what I understand, not unheard of in clinical settings.

     

    It isn't about questioning the validity of the technique/method per se (because I know this is widely accepted in analogous situations in other sectors), but rather about including some reference that someone has applied such framework/approach in the context of disease control, primarily as a part of literature review, since it appears that this isn't unheard of. I just haven't been able to come across any such paper/study on my own.

     




    Original Message