fixes

[jhnwu3]

This pull request adds new multimodal embedding runners for MIMIC-IV + CXR using both MLP and JambaEHR architectures, and brings several improvements and standardizations to the Mamba-based runner. The changes improve reproducibility, flexibility, and evaluation reporting for multimodal mortality prediction experiments.

New multimodal runner scripts:

• Added a new script multimodal_embedding_mlp_mimic4_cxr.py for running unified multimodal embeddings with an MLP backbone. This script supports EHR, notes, and X-ray metadata/negbio inputs, and includes argument parsing, dataset setup, model instantiation, training, evaluation, and benchmarking.
• Added a new script multimodal_embedding_jamba_mimic4_cxr.py for running unified multimodal embeddings with a JambaEHR (hybrid Transformer + Mamba) backbone. The script mirrors the structure and features of the MLP runner, supporting flexible configuration and benchmarking.

Enhancements and standardization to the Mamba runner:

Evaluation and metrics improvements:

• Changed the evaluation metrics in the Mamba runner (multimodal_embedding_mamba_mimic4_cxr.py) from only "accuracy" to include "pr_auc" and "roc_auc", aligning with the new runner scripts for better model assessment.
• Added code to print evaluation metrics after inference, providing clearer benchmarking output. [1] [2]

Argument parsing and reproducibility:

• Added a --condor flag to all runner scripts to facilitate setting the device to "cuda" (no index) for HTCondor GPU jobs. [1] [2] [3]
• Standardized the default batch size to 1 in the Mamba runner for consistency with the new scripts.

These updates provide a unified, extensible framework for benchmarking multimodal models on MIMIC-IV with CXR data, and facilitate reproducible experiments across different neural architectures.

References:

• [1] [2] [3] [4] [5] [6] [7]

[will-pang]

didn't test extensively, but I trust John haha