Table 3 Sequential models compared
From: Neural-signature methods for structured EHR prediction
Variants | Description |
|---|---|
\(S^2\) | Standard Signature with depth 2 |
\(LS^2\) | Log-signature—a condensed form of the signature |
\(LS^2 + LL\) | LS with lead-lag augmentation—extracts quadratic variation |
\(LS^2 + LL + ATI\) | \(LS + LL\) with time index—removes time reparameterisation invariance |
\(LS^2 + LL + ATD\) | \(LS + LL\) with time delta—accounts for non-uniform sampling rate |
\(LS^3 + LL + ATD\) | Increased truncation depth—more complex, sequential features |
BoW OH LR | Most basic model with one-hot bagging and logistic regression |
GRU | Baseline gated recurrent unit sequential model |
GRU + ATD | Takes into account time differences |