patient record model type breakdown
#work/patientsim
2024-02-06
# transformer
# RNNs
- DynEHR, metalearning with MAML atop an LSTM
- coupled rnns well-suited to dual data types, like image+poses
- EHR-M-GAN is based on an LSTM for its encoding of time series, which is based off DRAW.
- Merkelbach Paper just uses a GRU, pretty basic. One of the few that demonstrate the true autoencoder structure I’m looking to use any of this stuff for, though. Everything else just uses some sort of outcomes or next-event prediction or smth.
# Convolutional Neural Networks (CNNs)
- PatchMixer, very strange. It compares itself a lot with the transformer, though.
- dilated-convolutions-BC5HWKRF very weird