aaronwzl/gmdb_ten_disease_subset

GMDB Ten-Disease Subset (metadata + split)

A ten-syndrome subset of the GestaltMatcher Database (GMDB) v1.1.0, with a patient-level 70/30 train/test split.

This repo contains metadata only — no facial images. GMDB face images are controlled-access under a data use agreement; join on image_id against your own authorized GMDB copy (crop_path gives the aligned-crop filename).

Contents

1847 images from 1475 patients across 10 syndromes.

disease	images	patients	train	test	HPO available
Cornelia de Lange syndrome	447	361	312	135	39.8%
Williams-Beuren syndrome	273	249	191	82	36.6%
Noonan syndrome	207	182	145	62	54.6%
Kabuki syndrome	206	164	144	62	35.4%
KBG syndrome	179	134	125	54	68.7%
Angelman syndrome	160	125	111	49	13.8%
Rubinstein-Taybi syndrome	116	100	81	35	44.0%
Smith-Magenis syndrome	102	49	71	31	24.5%
Nicolaides-Baraitser syndrome	97	51	68	29	83.5%
22q11.2 deletion syndrome	60	60	41	19	86.7%

label is 0-9, assigned alphabetically over disease.

Syndrome selection

Nine classes map one-to-one onto a GMDB internal_syndrome_name. 22q11.2 deletion syndrome does not — GMDB splits it across three entries, which are merged here into a single class:

• CHROMOSOME 22q11.2 DELETION SYNDROME, DISTAL (32)
• DIGEORGE SYNDROME; DGS (25)
• VELOCARDIOFACIAL SYNDROME; VCFS (3)

Split methodology

• Grouping unit is patient_id, not image_id. Patients contribute up to 27 images each; an image-level split would leak. No patient appears in both splits.
• Stratified on disease × has_hpo, so both the class balance and the HPO availability rate are preserved. Every class lands within 30 ± 2% test, and train/test HPO coverage differs by ≤ 2.5 pp in every class.
• Patients are assigned greedily in descending image-count order, so the 70/30 ratio holds on images, not just on patient counts.
• Deterministic, seed=0.

The original GMDB benchmark split is preserved in gmdb_split (train/val/test/unassigned) for comparability with the GestaltMatcher literature. Note it is too thin on this subset for multimodal use — restricted to HPO-bearing images, Angelman has 0 val and 0 test.

Known caveat: HPO missingness is confounded with the label

Only 44.3% of images carry any HPO annotation, and the rate varies from 13.8% (Angelman) to 86.7% (22q11.2). "Is the phenotype field empty" is therefore itself a label signal. A model given both modalities can exploit the missingness pattern, and attribution methods will credit the absence of text rather than any phenotype content. The split above keeps this rate matched across train and test, but matching does not remove the confound — handle it explicitly (HPO-complete subset, or an explicit missingness indicator plus a has_hpo-only baseline).

HPO terms are patient-level constants: no patient in this subset carries more than one distinct HPO set.

Fields

field	notes
image_id, patient_id	GMDB identifiers; join keys
disease, label, split	this subset's class name, 0-9 index, and train/test assignment
present_features, absent_features	HPO ID lists (HP:…)
hpo_terms	human-readable names for present_features
n_hpo, has_hpo	count of present HPO terms, and whether it is non-zero
syndrome_id, syndrome_name, internal_syndrome_*, OMIM	original GMDB labels, pre-merge
gene_names, gene_entrez_ids, hgvs, test_type	molecular confirmation
ethnicity_*, age_year, age_month, age_note, gender	patient demographics
crop_path	aligned-crop filename in a GMDB image directory (image not included)
gmdb_split	original GMDB benchmark split

Contributor identifiers present in the raw GMDB export (username, email, author) and source identifiers (pmid, doi, filename) are not included.

Usage

from datasets import load_dataset

ds = load_dataset("aaronwzl/gmdb_ten_disease_subset")
print(ds["train"][0]["disease"], ds["train"][0]["hpo_terms"])

Citation

Please cite GMDB / GestaltMatcher:

Hsieh TC, Bar-Haim A, Moosa S, et al. GestaltMatcher facilitates rare disease matching using facial phenotype descriptors. Nature Genetics 54, 349-357 (2022).

Lesmann H, Hustinx A, Moosa S, et al. GestaltMatcher Database - a global reference for facial phenotypic variability in rare human diseases. medRxiv (2023).

Derived data is subject to the GMDB data use agreement.