-2
$\begingroup$

I am trying to figure out what would be the best way to learn patterns with a data-set that has temporal dependency between its features.

Let's say I want to predict whether a patient will suffer from a heart attack in the following minute by looking at his/her heart-rate, blood pressure and oxygen levels on each minute for the past five minutes.

Now each combination of the three features stated above has no predictive value on its own, but I assume that a temporal analysis might reveal some interesting predictive ability. For example, having a slow heart rate two minutes after heaving fast heart rate and low oxygen levels might point to an increase possibility of a heart attack.

What would be the best approach to tackle this kind of problem? Would a regular ML classification work in this case if I use each combination of index*time as a feature? (resulting in 3(index)*5(time-points) features).

Any suggestions/reading materials would be appreciated.

$\endgroup$
2
$\begingroup$

You might want to create new features from your data for different time instances. eq, heart rate at hour 1, heart rate at hour 2 ... These features once fed into your algorithm should answer your question.

| improve this answer | |
$\endgroup$
  • $\begingroup$ Thank you, but won't the inter-correlation between these features pose a problem? Which algorithm will treat the dependency between the features and not just the linear component of each feature? Am I making sense? $\endgroup$ – Zennie Nov 19 '17 at 6:12
  • $\begingroup$ I dont think there would be any correlation among these new features as the values at different times won't have any relation with each other and that's the reason we are consideration this data. Say, a patient has heart rate x in first hour and y in second hour. These values could be same and could be different but for various patients, this has to be non-correlated. I did work on a similar problem recently and it worked pretty decently. $\endgroup$ – emudria Nov 19 '17 at 10:54

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.