I am an undergraduate with near-infinite passion to the theoretical machine learning and ML applications. Inspired from my challenging mental disorders, I am really interested in building a diagnostic system for the psychiatric disorders (with focus on thought-driven disorders) that could diagnose a patient with certain mental disorders, and even predict what he/she might have in future. I have a couple of questions:

1) Beside knowledge about the machine learning and implementation of its algorithms, what do I need to build diagnostic system? Do I need to study recommender system too?

2) As you might know, we do not have big data for psychiatric disorders due to different regulations for privacy of patients. Is there a technique in machine learning designed for dealing with "small data" (perhaps an ensemble learning like a random forest)?

3) Following from the question above, is there a technique that could create simulations of what patients' record might look like (i.e. create new but hypothesized datum about patients to build small data into big data)?

4) Is there a way to translate ordinal data into cardinal data? As you might know, the existing diagnostic methods for mental disorders are in ordinal form (i.e. your symptom is from 1-10)? Due to different diagnostic methods, I think learning by rank will not be useful...

Thank you very much for your time, and I am sorry for this long post...

  • 2
    $\begingroup$ What kind of data do you think you would actually be using to predict the presence of a disorder? Mental disorders are generally defined (in the DSM) in terms of fairly short lists of symptoms, with clear criteria for which symptoms are required for a diagnosis. ...but the symptoms are not really "objective", and generally you need to be a trained psychiatrist/psychologist to properly interpret them. $\endgroup$
    – Marius
    May 9, 2017 at 5:09
  • $\begingroup$ The problem with extrapolating from a small data set is that you risk not introducing anything new to the mix for the model to learn, in which case the additional data is redundant. If you try to subvert that by introducing small random variations, there is a risk that it stops simulating the real world and a model trained on that may not perform well on real patients. $\endgroup$
    – Antimony
    May 11, 2017 at 19:00


Your Answer

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