# Best Way to tackle to time series classification problem?

I have a dataset where the input is a dataset for ICU patients where each ICU stay has 40 features (20 vitals, 20 lab values) and multiple time steps (the stays' length is between 6 and 19-time steps). The output is the 20 lab values (represented as binary numbers where 1 means the lab value is in range and 0 is out of range). The task is to predict the future output for an input ICU stay. I have tried to make it a sequence to sequence problem where the output is a sequence shifted by a one-time step. However, the results were not good enough. The question is there a way to better tackle this problem? (like maybe windowing ?)

So, if you have 20-element array of binary numbers as output, you might find out there is no two ouput exactly the same or there might be only a few observations with exactly similar output (because you can have $$2^{20}$$ combinations!). So, let's say if you have N observations, and you may see N distinct output array as well. Therefore, you may first cluster the labels into k groups and assign new labels to each group (the label of group i-th is i). Then, find the closest distance between your query and each group, so you can then find the appropriate class. Then, you can predict the ouput in array format (e.g. let's say you realize your query belongs to a group of 20 members). So, you can say, the output is expected value of output arrays in element-wise manner. So, e.g. your first element of 20-element output array might be 0.9. Or, you can be more precise and after finding the appropriate group, define a weight based on the distance between your query and members and find the weighted average of output arrays of that group!