RNNs for classifying irregularly sampled signals


I’m tackling a problem in which a time-varying measurement (a signal) must be classified in one of two groups. Since different signals have different lengths, using a recurrent neural network seemed like a good idea. However, the signals are not regularly sampled nor aligned, that meaning the first signal might have 2 measurements at times 2 and 3, whereas the second signal might have 3 measurements at times 6, 10, 12. I tried training an RNN in which the time is yet another feature (2 features, overall) but that yields worse results than using only the measurement (also, not sure if including the time as a separate features makes sense). So, my question is: how can I pass to the RNN the time stamp associated with each sample in a meaningful way?