Clinical trajectories and genetic architecture across the neurological–psychiatric boundary

Disease-trajectory similarity computed from Delphi-2M embeddings recovers recognizable clinical relationships: major depressive disorder and bipolar disorder were the most similar psychiatric pair (cosine similarity 0.93), anxiety and essential tremor were close despite sitting on opposite sides of the boundary (0.91), and PTSD and Alzheimer's disease were among the least similar pairs (0.58), consistent with their very different ages of onset. No comparable measure of clinical similarity was previously available at scale; this work uses embeddings from an electronic-health-record transformer directly as a quantitative measure of disease trajectory similarity. Similarities are cosine similarities between embeddings of a model retrained on records from about 400,000 UK Biobank participants, and they align with recognized clinical relationships.

Clinical and genetic similarity correlated moderately across the 171 disorder pairs (Mantel test r = 0.33, p < 1×10⁻⁴), and the correlation remained significant for other measures of genetic similarity. This is the first direct comparison, across the same 19 disorders, of clinical trajectory representations learned from health records with genome-wide genetic correlations. The Mantel test assessed significance with 10,000 simultaneous row-and-column permutations; genetic correlations were taken from Smeland et al. LDSC estimates and were available for all 19 disorders.

Within-category similarity exceeded cross-category similarity in both measures (p = 1.9×10⁻³ for clinical and p = 1.1×10⁻⁵ for genetic similarity), and psychiatric pairs were on average more similar to one another (mean cosine 0.84 versus 0.80; mean rg 0.32 versus 0.11), making neurological disorders the more heterogeneous group. The boundary was operationalized as average within-category similarity minus average cross-category similarity and computed separately for clinical and genetic measures, making crossing a comparable quantity. An exact permutation test covered all 92,378 ways of assigning nine of the 19 disorders to the neurological category.

Sixteen of the nineteen disorders respected the category boundary on both measures, and the two measures made the same exceptions: multiple sclerosis, migraine, and essential tremor sat closer to psychiatric disorders in both clinical trajectory and genetic similarity, an agreement expected by chance in about one in 969 cases. Cross-boundary pleiotropy had mainly been reported at the genetic level; here clinical trajectories learned from health records break the boundary in the same places. The crossing held in every leave-one-disorder-out iteration and in ten retrainings of Delphi-2M with different random seeds, and the three disorders have between 1,968 and 20,508 UK Biobank cases.

The work extends the question of whether the boundary is real from the genetic level to the level of clinical trajectories, and it offers a reusable path: comparing representations trained on health records against external biological signals to distinguish learning disease processes from learning the process of care. It mainly serves researchers who want to test whether electronic-health-record embeddings carry biological structure, along with clinical and research teams interested in more integrated neurological–psychiatric care; the conclusions apply to settings with ICD-10 three-character-style coding, similar specialty divisions, and UK Biobank-like cohorts, and the correspondence with genetic architecture applies to the 19 analyzed disorders and European-ancestry common-variant estimates.

How much of the proximity in the embeddings reflects disease processes versus care pathways remains an open question: genetic liability can influence whether and when people seek care, and the two measures share the UK Biobank cohort, so they are not fully independent. Genetic correlation estimates depend on how cases are diagnosed and can be inflated by misclassification, and they rest on common variants in European-ancestry individuals; the disorder set is limited by GWAS summary-statistic availability. Genetic overlap measured by MiXeR could be estimated stably for only 13 disorders, and the set of disorders crossing the genetic boundary under MiXeR was partly different, although migraine crossed on every measure. In addition, this assessment is based on the loaded main text and figure captions; details of Figures S1–S6 and Tables S1–S2 were not included in the loaded text, so specific values from those supplementary results remain open to checking.