Unified Information Bottleneck Framework for Multimodal Biomedical ML

Code and aggregate results accompanying:

Dong, L. A Unified Information Bottleneck Framework for Multimodal Biomedical Machine Learning. Entropy 2026, 28(4), 445. https://doi.org/10.3390/e28040445

The paper develops an information-theoretic framework for multimodal biomedical machine learning, formulating multimodal representation learning through the information bottleneck principle and providing tools for: mutual-information decomposition across modalities, redundancy/synergy quantification, fusion-collapse diagnostics, missing-modality robustness as information consistency, longitudinal modeling via transfer entropy, and uncertainty / calibration / OOD detection. Empirical case studies span three datasets (TCGA-BRCA, TCGA-GBMLGG, OASIS-2).

This repository contains:

• experiments/ — Python source for all empirical case studies in Section 9
• experiments/results/ — aggregate result JSON files reported in the paper
• environment.yml, requirements.txt — reproducible environment specs
• CITATION.cff — machine-readable citation

This repository does not redistribute any patient-derived data.

---

Quick start

git clone https://github.com/ProfessorDong/unified-ib-multimodal-biomed.git
cd unified-ib-multimodal-biomed
conda env create -f environment.yml
conda activate nih_research
# obtain data per the Data Availability section below, then:
cd experiments && python3 run_all.py

See experiments/README.md for per-script details and expected paths.

---

Data Availability

Reproduced verbatim from the paper:

The empirical studies in Section 9 use three publicly available datasets: (1) the TCGA-BRCA multi-omics dataset as preprocessed by Wang et al. (available at https://github.com/txWang/MOGONET, accessed on 15 February 2026); (2) the TCGA-GBMLGG clinical and genomic dataset from Chen et al. (available at https://github.com/mahmoodlab/MCAT, accessed on 15 February 2026); and (3) the OASIS-2 longitudinal Alzheimer's MRI dataset (available at https://www.kaggle.com/datasets/jboysen/mri-and-alzheimers, accessed on 15 February 2026). No new primary data were generated.

OASIS-2 is governed by a Data Use Agreement; users must register and agree to OASIS terms before downloading. After download, place files so the paths in experiments/README.md resolve.

---

Selected results

Quantity	Value
mRNA mutual information, BRCA	0.878 ± 0.044 nats (64.9% of H(Y))
Synergy S, BRCA same-level fusion	≈ −0.63 (negative across 5/5 folds)
Synergy S, GBMLGG cross-level	≈ −0.16 (near zero)
Best unimodal AUC, BRCA	0.928 ± 0.013
VMIB AUC, BRCA	0.915 ± 0.016
Missing-modality worst case (Standard / Dropout)	0.668 / 0.796
Fusion balance index (Standard / Dropout)	0.25 / 0.62
OOD entropy ratio	2.0×
Selective prediction at 50% coverage	0.939 accuracy (vs 0.787 base)
Sequential prediction gain, OASIS-2	+0.022 AUC
Consistency-model degradation reduction	37%

Full numbers, including confidence intervals, are in experiments/results/*.json and the paper.

---

How to cite

Preferred (paper):

@article{e28040445,
  author  = {Dong, Liang},
  title   = {A Unified Information Bottleneck Framework for Multimodal Biomedical Machine Learning},
  journal = {Entropy},
  volume  = {28},
  year    = {2026},
  number  = {4},
  pages   = {445},
  doi     = {10.3390/e28040445},
  url     = {https://www.mdpi.com/1099-4300/28/4/445},
  issn    = {1099-4300}
}

Software (Zenodo). Use the concept DOI to cite the software in general (always resolves to the latest archived version); use the version DOI when you need to pin a specific release for reproducibility.

@software{dong_2026_zenodo_concept,
  author    = {Dong, Liang},
  title     = {Unified Information Bottleneck Framework for Multimodal Biomedical ML — code and results},
  year      = {2026},
  publisher = {Zenodo},
  doi       = {10.5281/zenodo.19802841},
  url       = {https://doi.org/10.5281/zenodo.19802841},
  note      = {Concept DOI; resolves to the latest archived release}
}

@software{dong_2026_zenodo_v1_0_1,
  author    = {Dong, Liang},
  title     = {Unified Information Bottleneck Framework for Multimodal Biomedical ML — code and results},
  year      = {2026},
  publisher = {Zenodo},
  version   = {v1.0.1},
  doi       = {10.5281/zenodo.19802842},
  url       = {https://doi.org/10.5281/zenodo.19802842}
}

---

Funding

This research was funded by the National Cancer Institute (NCI) of the National Institutes of Health (NIH) under Grant R01 CA309499.

License

Code is released under the MIT License. The published paper is open access under CC BY 4.0 at the MDPI Entropy DOI above.

Contact

Liang Dong — Baylor University / UT Southwestern Medical Center. For issues with the code, please open a GitHub issue.