AdaCare model for explainable clinical health status representation learning. Supports both code-based features (with embedding) and timeseries features (direct).
Usage
AdaCare(
dataset = NULL,
embedding_dim = 128,
hidden_dim = 128,
kernel_size = 2,
kernel_num = 64,
r_v = 4,
r_c = 4,
activation = "sigmoid",
rnn_type = "gru",
dropout = 0.5
)Arguments
- dataset
A
SampleDatasetobject providing input/output schema- embedding_dim
Integer, embedding dimension for code features. Default 128
Integer, hidden dimension for RNN. Default 128
- kernel_size
Integer, kernel size for convolutions. Default 2
- kernel_num
Integer, number of kernels per scale. Default 64
- r_v
Integer, reduction rate for input recalibration. Default 4
- r_c
Integer, reduction rate for conv recalibration. Default 4
- activation
Character, activation function. Default "sigmoid"
- rnn_type
Character, RNN type ("gru" or "lstm"). Default "gru"
- dropout
Numeric, dropout rate. Default 0.5
Details
Paper: Ma et al. "AdaCare: Explainable clinical health status representation learning via scale-adaptive feature extraction and recalibration." AAAI 2020.
This model automatically detects feature types and processes them appropriately:
SequenceProcessor: Code-based features → Embedding → AdaCareLayer
TimeseriesProcessor: Numerical features → AdaCareLayer (direct)
Returns attention weights for interpretability.