Add/multimodal rnn + soft fix mask absolute value bug

[jhnwu3]

This pull request modernizes and standardizes several PyHealth example scripts for clinical prediction tasks using the new BaseDataset and BaseTask classes. It introduces improved workflows for drug recommendation, mortality prediction, readmission prediction, and length of stay prediction on eICU and MIMIC-IV datasets, as well as demonstrates the new MultimodalRNN model. The updates focus on clarity, maintainability, and leveraging the latest PyHealth features.

Major updates to example scripts:

Modernized eICU examples:

• Added new example scripts for drug recommendation, mortality prediction, readmission prediction, and length of stay prediction on eICU, each using the new BaseDataset-based eICUDataset and the corresponding BaseTask classes (DrugRecommendationEICU, MortalityPredictionEICU, ReadmissionPredictionEICU, LengthOfStayPredictioneICU). These examples demonstrate the standardized PyHealth workflow, including dataset loading, task setting, splitting, model definition, training, and evaluation. [1] [2] [3] [4]

Multimodal model demonstration:

• Added a new example script for in-hospital mortality prediction on MIMIC-IV using the MultimodalRNN model, showcasing handling of both sequential and non-sequential features, model initialization, training, evaluation, and prediction output inspection.

Documentation improvements:

• Updated API documentation to include the new MultimodalRNN class, ensuring users can easily find information about its usage and features.

Deprecation of legacy workflow:

• Removed the legacy drug recommendation example script that used custom task functions and the old workflow, replacing it with the new standardized version.

---

References:
[1] [2] [3] [4] [5] [6] [7]

[jhnwu3]

nvm need to reset some details as it includes files from a separate branch