Is it OK to re-sample the same patients at multiple points in time and then use the data for classification?

Question

I'm trying to predict which patients will have a bad outcome (BO) within the next 2 weeks based on medical readings (plus some behavioral data) from the last 6 weeks. Total data timespan is 8 weeks, i.e. each patient has 6 full weeks of data in the feature space, and I'm trying to predict BO (yes/no) within the next 2 weeks.

Re-sample at different times-- To prevent the model from being biased to a specific point in time, and to increase the number of BO cases for predictive modeling (there's only ~40 BO cases a week), I'm thinking taking many points in time as "as-of dates", and for each patient at each as-of date I create features looking 6 weeks back and BO(yes/no) looking 2 weeks ahead. For eg. John Doe was sampled at 10 as-of dates and he becomes 10 rows in the data set.

My question is this -- I understand that the 10 John Doe rows are not independent from one another, so I'll avoid using models that require independent assumption such as logistic regression. Can I use ensemble tree models to predict BO (such as Gradient-boosted tree, XGBoost, Random Forest, etc.)?

Additional detail (if that helps)-- Samples from one as-of date to the next as-of date only have ~85% overlapping patients. There's new patients entering each sample, and there's patient rolling out (released or BO).

Any advice will be much appreciated!

Answer 1

Yes, you can use multiple observations of the same patients taken at different times. I recommend consulting with a statistician for details on how to do time series regression. Using XGBoost or fancy ML tree algorithms won’t help here, though, and chances are they won’t be all that useful. All of these algorithms are designed for fast nonparametrics with large datasets, where computation time is an issue and variance isn’t. Instead, I would recommend a much simpler model like ARMA with random effects, or a simple Markov model.