Disease Embedding Space
Summary¶
Jia, Li, Zhong, et al. (2023) apply a word2vec-style embedding to 547 broadly-defined diseases, trained on chronologically-ordered diagnosis histories from over 151 million patients in the Merative MarketScan claims database, treating each patient's disease history as a "sentence" and each diagnosis as a "word." The result is a 20-dimensional continuous space in which diseases with similar etiology and clinical manifestation cluster together. Because each of the 20 dimensions behaves like a continuous quantitative trait, and each patient occupies a point in the space (a weighted mean of their diagnosed diseases' coordinates), the authors performed genome-wide association studies directly on these 20 dimensions in the UK Biobank, discovering 116 genome-wide-significant associations across 108 loci (40 previously unreported), independently replicated in Japanese (BioBank Japan) and US (BioVU) cohorts.
Method¶
- Embedding training: a shallow neural network (word2vec-style skip-gram/CBOW) is trained to predict a randomly masked diagnosis from its surrounding diagnosis context within a patient's chronological health record, over 547 disease categories (each an aggregation of non-overlapping ICD codes) from >151 million MarketScan patients.
- Dimensionality and orientation: the resulting embedding is rotated via PCA so dimension 1 captures the greatest variance, through dimension 20.
- Interpretation: cosine similarity between disease vectors recovers intuitive clinical relationships (e.g., CNS degenerative diseases' most distant "antipode" is fasciitis, an unrelated athletic injury); vector arithmetic recovers approximate disease analogies (e.g., chondrocalcinosis + connective tissue infection ≈ septic arthritis). Each of the 20 dimensions was manually labeled by testing which of 21 broad disease categories separate significantly along it (e.g., dimension 8: "ophthalmological (−) to CNS (+)"; dimension 2: "neuropsychiatric (−) to infectious (+)").
- Disease constellations: singular value decomposition identifies 10 maximally distinct disease clusters ("constellations") anchoring the embedding space.
- Patient-level quantitative traits: a patient's health state is represented as the weighted mean of their diagnosed diseases' 20-dimensional coordinates, yielding 20 continuous phenotypes usable directly in standard GWAS and family-heritability pipelines.
Genetic Findings¶
- Heritability: family-based heritability (128,989 nuclear families, 481,657 individuals) and UK Biobank/BioBank Japan GWAS-based heritability were estimated for all 20 dimensions; the two heritability estimates correlated positively (r=0.224, 95% CI [0.085, 0.355]) despite being derived from independent methods and largely non-overlapping cohorts. Individual-dimension heritability did not exceed 0.2, but joint co-heritability across all 20 dimensions reached 0.187 due to substantial cross-dimension genetic correlation. Dimension 19 ("immune (−) to neuropsychiatric (+)") had the largest family heritability and shared-environment contribution; dimension 15 ("metabolic (−) to infectious (+)") had the lowest.
- GWAS in UK Biobank: 116 genome-wide-significant association signals across 108 unique loci (Bonferroni-corrected), 40 of which were not previously reported in any GWAS. For example, dimension 8 associated with six loci near immune-transcription-factor genes (BACH2, GTF3AP1, STAT6, EMSY, SMAD3), plausibly linking migraine and inflammatory bowel syndrome through shared IL-4-mediated immune signaling.
- Replication: of the 116 associations, 100 were testable in BioBank Japan (16 involved loci not genotyped/imputed there) — 87 replicated in the same effect direction. In BioVU (Vanderbilt), 87 of 116 associations replicated in the same effect direction.
Significance¶
This represents an NLP-inspired alternative to more common multimodal deep-learning approaches (e.g., RisQ, Delphi-2M) for capturing the shared structure of human disease — using a shallow, interpretable word2vec-style model on diagnosis codes alone, rather than a large transformer over multimodal clinical data, and explicitly connecting the resulting continuous disease representation back to GWAS-discoverable genetic loci.
Availability¶
Reproducible code and data via a Code Ocean capsule: doi.org/10.24433/CO.0096653.v1. Underlying MarketScan data requires a commercial license; UK Biobank, BioVU, and BioBank Japan data require standard controlled-access applications.
See Also¶
- Multimodal Disease Risk Prediction — related paradigm of modeling shared structure across diseases, using deep sequence models rather than a word-embedding approach.