This work builds upon the HAIM framework originally developed by Hakima Laribi in the MEDomicsLab/HAIM repository.
Data access: The MIMIC-IV multimodal dataset requires credentialed access via PhysioNet.
| File | Description |
|---|---|
src/data/dataset_adapted_poc.py |
Adapted HAIMDataset with flexible modality selection and float32 compression for MEDomics integration |
notebooks_poc/enlarged_cardiomediastinum.ipynb |
Compression validation experiment — Enlarged Cardiomediastinum task |
notebooks_poc/consolidation.ipynb |
Compression validation experiment — Consolidation task |
These notebooks validate that compressing HAIM features from float64
to float32 (scaled by 100,000) preserves model performance, a
necessary step for integrating the dataset into the
MEDomics platform.
AUC values obtained for both clinical tasks. HAIM original values are visual estimates extracted from the supplementary material of the original study. Code values are computed using the non-compressed and compressed versions of the dataset.
| Task | HAIM original (estimate) | Code (non-compressed) | Code (compressed) |
|---|---|---|---|
| Enlarged Cardiomediastinum | ~0.80 | 0.81 ± 0.03 | 0.82 ± 0.02 |
| Consolidation | ~0.85 | 0.88 ± 0.02 | 0.87 ± 0.02 |
The small difference between compressed and non-compressed results (≤ 0.01)
confirms that float64 → float32 compression preserves model performance.