How do I deal with data that has only limited target values?

Question

I'm currently working on a small project using the D1NAMO dataset (1). I want to predict the glucose level (that is given in the dataset) based on several features: accelerometer data, heartbeat (ECG) data, breathing data and some other features. I have several users (006, 007, 008 etc) that all have a few days of data.

The problem is that the target value (glucose) is measured only once in 5 minutes, whereas the other measurements are done every few milliseconds. That means that in my data I have many rows without the target value. For example, for one user (008) I have 214 rows where I have a glucose level and around 60.000 rows where glucose is NaN. I want to use a model without the notion of time, so that could be anything such as FFNN, CNN, Decision Trees (XGBoost), ensembles etc.

Snapshot of my data:

My questions:

1. How should I deal with this problem where I have only few target values? Should I summarize all values in a window of 5 minutes to match with the target value?
2. Should I make a model on the instance level or population level?

Thanks in advance for any replies.

(1) https://www-sciencedirect-com.vu-nl.idm.oclc.org/science/article/pii/S2352914818301059

Answer 1

I would try two different approaches:

• interpolate the missing values on a user level.
• work with the sunset of rows for which we actually have the glucose level.

Then, I would compare the test accuracy of the model built with both methods. Remember that your test set has to be composed of rows for which you have the glucose level - you cannot build it with interpolated data, that would be cheating!