I have the following time-series features: Diastolic Blood Pressure, Systolic Blood Pressure, Heart Rate, RR variability and Arterial Blood Pressure. Each of these clinical parameters was measured for 900 seconds during a surgical procedure and after the surgery, the patient was assessed for acute kidney injury: 1(yes) or 0(no).

My training data kind of looks like this: (see below for screenshot)

Patient 1 Time(s) Features AKI

Patient 2 Time(s) Features AKI

and so on.

What approach would I take to utilize this data for the binary classification task?

Sample view of Training Data

  • $\begingroup$ try lstm for your ptoblem $\endgroup$ – Andreas Look Mar 20 '19 at 21:23
  • $\begingroup$ Is there any machine-learning approaches to this? Not too familiar with RNNs. $\endgroup$ – John Spanos Mar 20 '19 at 21:41
  • $\begingroup$ Somehow related question / answer datascience.stackexchange.com/a/25518/29781 $\endgroup$ – aivanov Mar 22 '19 at 13:52

I would plot the measurements time $t$ and the corresponding measurement 5 to 10 samples for each category. Try to detect some patterns. Possible patterns are the trend (Is the cure growing? Is it linearly? Exponentially?), frequencies of oscillations (Does one category have oscillations with higher magnitude or frequencies? You can use Fast Fourier Transform for this) self similarity of signal (autocorrelation) Then look at means, median, standard deviation, skewness and kurtosis of your signals.

After having extracted all these features I would try to calculate the correlations of your features with the target variable. Then you can eliminate variables that are not very highly correlated with your target variable. In the next step, I would look at the correlations between your features and eliminate the variables that are highly correlated by eliminating the one variable which less correlated with the target variable. Then I would use some classical binary classifiers like discriminant analysis or logistic regression.

If you see that this method will not lead to sufficient results then you should try more sophisticated methods like neural networks/decision trees for the features that you extracted.

| improve this answer | |
  • $\begingroup$ Hi, thanks for the answer. To clarify, the target variable isn't numerical. It is a categorical value: either the patient has aki or not. $\endgroup$ – John Spanos Mar 20 '19 at 23:08
  • $\begingroup$ My explanation is still applicable for these cases. The correlation will be the point-biserial correlation. Everything else can be applied as proposed. $\endgroup$ – MachineLearner Mar 21 '19 at 7:10

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.