Age-dependent topic modelling of comorbidities in UK Biobank identifies disease subtypes with differential genetic risk

Abstract Longitudinal data from electronic health records (EHR) has immense potential to improve clinical diagnoses and personalised medicine, motivating efforts to identify disease subtypes from age-dependent patient comorbidity information. We introduce an age-dependent topic modelling (ATM) method that provides a low-rank representation of longitudinal records of hundreds of distinct diseases in large EHR data sets. The model learns, and assigns to each individual, topic weights for several disease topics, each of which reflects a set of diseases that tend to co-occur as a function of age. Simulations show that ATM attains high accuracy in distinguishing distinct age-dependent comorbidity profiles. We applied ATM to 282,957 UK Biobank samples, analysing 1,726,144 disease diagnoses spanning 348 diseases with ≥1,000 incidences. We inferred 10 disease topics optimising model fit. We identified 52 diseases with heterogeneous comorbidity profiles (≥500 incidences assigned to each of ≥2 topics), including breast cancer, type 2 diabetes (T2D), hypertension, and hypercholesterolemia; for most of these diseases, topic assignments were highly age-dependent, suggesting differences in disease aetiology for early-onset vs. late-onset disease. We defined subtypes of the 52 heterogeneous diseases based on the topic assignments, and compared genetic risk across subtypes using polygenic risk scores (PRS). We identified 18 disease subtypes whose PRS differed significantly from other subtypes of the same disease, including a subtype of T2D characterised by cardiovascular comorbidities and a subtype of asthma characterised by dermatological comorbidities. We further identified specific SNPs underlying these differences. For example, the T2D-associated SNP rs1063192 in the CDKN2B locus has a higher odds ratio in the top quartile of cardiovascular topic weight (1.19±0.02) than in the bottom quartile (1.08±0.02) (P=4×10-5 for difference). In conclusion, ATM identifies disease subtypes with differential genome-wide and locus-specific genetic risk profiles.