I have an EHR data source which has info on

a) Patient visit records (Inpatient, outpatient, Emergency etc) and why did he visit hospital (diagnosis codes attached to each visit)

b) Patients drugs data

c) Patients lab test data

d) Patient diagnosis history

e) Patient demographics data

If my objective had just been to predict when will we visit hospital next, I can use Time to event analysis or survival analysis etc.

But, I would also like to know why will he visit hospital next or what might be the reason of his next hospital visit?

While predicting "when" seems to be doable, can experts here guide me on how can I find why/what will be the reason for his next hospital visit?

  • $\begingroup$ I have doubts about the use of the task: aren't most visits scheduled appointments for consultations, treatment or tests? If yes then the model would just attempt to reproduce the regularity of the appointments, which is an information that the doctors/hospital already have. I would also guess that very often people have regular visits for the same condition, so predicting the "why" could probably be done just by picking the most frequent diagnosis code in the recent past. For unexpected visits like ER it's mostly unpredictable anyway (and most patients probably wouldn't have an EHR). $\endgroup$ – Erwan Feb 4 at 23:39
  • $\begingroup$ So I'm not convinced but technically you could try to use all these features to predict the when and why. It would be a sequence task, I know only the "old" CRF model, but I think there are more recent DL approaches which would probably work better. $\endgroup$ – Erwan Feb 4 at 23:46
  • $\begingroup$ @Erwan - Let's say for ex: we analyze hypertension patients. And hypertension can have a list of complications. I would like to what will be the reason (which of those hypertension complications) may lead him to visit next. $\endgroup$ – The Great Feb 5 at 0:46
  • $\begingroup$ @Erwan - May I kindly check whether you can share any similar paper where I can read about this? I am currently unable to visualize the logic. $\endgroup$ – The Great Feb 5 at 8:14
  • $\begingroup$ Ok, I thought your dataset was for any random patient going to the hospital, it makes more sense for following patients with a specific condition indeed. In this case the problem may be framed as predicting the probability of particular complications given the history/data of the patient. In the simplest form it's a classification problem where the class is one of the possible problems which can happen with hypertension. In a more advanced form I guess it can be seen as some kind of sequence classification problem, where each step in the sequence consists of all the features at this point ... $\endgroup$ – Erwan Feb 5 at 11:55

Why is potentially hidden in your feature explanations.

Given that your Features give some indication on why, like Patient demographics data, or some other Features that you can include here, you can use them to answer the why.

Like this using shapley and or eli5 you can integrate it with your Standard predictor classes and for the explanations with user friendly API.

BUT i have to prefice that even tough it might be hidden it must not be at the same time, since the predictiveness of time might have Nothing to do real why you are Looking for.

  • $\begingroup$ Thanks for your response. Upvoted $\endgroup$ – The Great Feb 16 at 14:25

Your Answer

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

Not the answer you're looking for? Browse other questions tagged or ask your own question.