Contributor: Nikhita Shanker (nikhita2@illinois.edu)

Contribution Type: Model

Original Paper: "Robust Mortality Prediction in the Intensive Care Unit using Temporal Difference Learning" (Frost et al.)
Link to the Original Paper: https://arxiv.org/pdf/2411.04285

Description: This PR adds the CNNLSTMPredictor model to PyHealth, implementing the CNN-LSTM model from the paper. The model processes medical codes through:

1. Embedding layer - Maps discrete medical codes to dense vectors
2. CNN encoder - Uses (Conv1d -> BatchNorm -> ReLU -> MaxPool) for capturing local patterns
3. LSTM encoder - For sequential dependency modeling
4. Dense decoder - Uses (BatchNorm -> Linear -> ReLU -> Dropout -> BatchNorm -> Linear) and produces the final prediction

• The model inherits from pyhealth.models.BaseModel, and implements the forward() method

Replication Aspect of the Paper: This PR directly implements the CNN-LSTM model from the replicated paper (Frost et al.).

• The architecture in cnn_lstm.py uses the original paper's CNN-LSTM design
• The ablation/extension extends the paper through hyperparameter variation (learning rate, hidden dimension, dropout, batch size), whereas the original paper uses fixed default values for these hyperparameters.

Files to Review:

• pyhealth/models/cnn_lstm.py: Model implementation (CNNLSTMPredictor)
• pyhealth/models/init.py: Adds CNNLSTMPredictor to imports
• tests/core/test_cnn_lstm.py: Unit tests with synthetic data
• examples/mimic4_icu_mortality_cnn_lstm.py: Ablation/extension script with learning rate, hidden dim, dropout, and batch size sweeps
• docs/api/models/pyhealth.models.cnn_lstm.rst: API documentation RST file
• docs/api/models.rst: Updated to add cnn_lstm