Across three major biobanks, the framework traced how interconnected disease risks change over the life course and exposed differences hidden within conventional diagnoses.
Study: A Bayesian framework for longitudinal EHR and genetic discovery. Image Credit: Francois Poirier / Shutterstock
In a recent study published in Nature, researchers introduced ALADYNOULLI, a Bayesian generative model that reveals underlying disease signatures by integrating electronic health records (EHRs) and genetic data.
The framework showed stronger discrimination of short- and long-term risk than established clinical risk scores in the evaluated United Kingdom Biobank comparisons. The model could also provide potentially useful clinical information by predicting disease-level phenotypes rather than individual diagnostic codes. Following prospective validation, the framework could help identify individuals who may benefit from closer preventive assessment.
People differ in their likelihood of developing diseases due to genetic, medical, and environmental factors. Disease risk may also vary by age, environmental exposures, and lifestyle factors. EHRs are a valuable source of longitudinal information that, when linked to individual genetic susceptibilities, could help dissect complex population-level disease trajectories. Comprehensive risk prediction models that consider diseases at a broader level and incorporate germline genetics may support disease prevention, early diagnosis, and prompt treatment.
About the Study
In the present study, researchers presented a Bayesian generative framework that modeled age and EHR diagnoses along with polygenic risk scores (PRS) to reveal temporal patterns in disease risk and disease progression among diagnostic subgroups. The Bayesian framework combined signature-specific probabilities to predict future disease onset while accounting for multiple coexisting disease processes and persistent chronic conditions.
The team applied ALADYNOULLI to three different datasets to assess cross-cohort reproducibility, although risk prediction performance was evaluated in the United Kingdom Biobank. The model integrated information from the All of Us (AoU), Mass General Brigham (MGB), and United Kingdom Biobank (UKB). The datasets comprised records from more than 683,000 participants covering 348 PheCode-defined disease phenotypes. The longitudinal records spanned up to 52 years, although follow-up duration varied between individuals and cohorts. To investigate genetic influences on the signatures, the researchers performed signature-based genome-wide association studies (GWAS) and rare variant association studies (RVAS), and compared selected findings with conventional single-disease analyses.
The team used PheCodes, which group related ICD-10 diagnostic codes into disease-level phenotypes, for risk prediction. They used inverse probability weighting (IPW) to address potential UKB participation bias while preserving core disease-signature relationships. The researchers compared model performance with established clinical risk scores, including the Pooled Cohort Equation (PCE) and PREVENT for atherosclerotic cardiovascular disease, and the Gail model for breast cancer, over one-year short-term and 10-year long-term follow-up periods. They also assessed how the framework performed relative to Delphi-2M, an AI model for predicting individual diagnostic codes.
Results
With the number of latent components set to 21, comprising 20 disease signatures and one low-incidence reference signature, the model showed high cross-cohort preservation of signature composition (median 80%) and revealed distinct patient subgroups within broader diagnostic categories. The signatures reflected known biological characteristics of the diseases. Individuals carrying familial hypercholesterolaemia-associated variants showed enrichment of the cardiovascular signature. Individuals with clonal haematopoiesis of indeterminate potential showed stronger inflammatory signatures. Rare variant burdens in low-density lipoprotein receptor (LDLR), titin (TTN), and BRCA2 aligned with their respective disease-associated patterns. Disease signatures were also strongly associated with inherited genetic risk. For example, higher PRS for coronary artery disease or elevated LDL corresponded to a stronger cardiovascular disease signature.
Within the non-ischaemic cardiovascular signature, the modeled probabilities of atrial fibrillation and heart failure increased progressively after the age of 55 years. South Asian genetic ancestry was associated with greater cardiovascular signature loading, which peaked at approximately 50–60 years and remained elevated in older age. Within the malignancy signature, the modeled probability of metastatic disease increased sharply between 60–75 years. The team found similar age-related disease patterns across the datasets.
The model also learned disease progression patterns from patient data that were consistent with established medical knowledge. ALADYNOULLI captured clinically expected sequences in which hypercholesterolaemia preceded myocardial infarction and primary cancers preceded metastatic disease. The cardiovascular signature loading rose more rapidly before myocardial infarction in early-onset cases than in late-onset cases, potentially reflecting different biological mechanisms despite the same diagnosis. Descriptive clustering also identified subgroups of patients with depression or breast cancer who had greater loadings for inflammatory and metabolic disease signatures.
The signature-based GWAS analyses identified 151 genome-wide significant loci, including some cardiovascular associations that were not identified as lead loci in the researchers' constituent cardiovascular GWAS. ALADYNOULLI achieved higher areas under the curve (AUCs) for short- and long-term disease prediction than established clinical risk scores in the evaluated UKB comparisons, although the authors did not conduct separate hypothesis tests for every pairwise comparison. The PheCode predictions also complemented ICD code-level models by providing disease-level risk estimates, although the model's effects on clinical decision-making and treatment prioritization were not tested.
Conclusions
Based on the findings, ALADYNOULLI captures shared disease patterns and achieves stronger risk discrimination than existing clinical scores in the evaluated comparisons. The model generated highly stable disease signatures across cohorts that were consistent with established clinical phenotypes in the examples directly examined.
By integrating longitudinal EHR and genetic data, the framework estimates how disease risk changes across the lifespan, including conditions with comparatively limited data. However, incomplete EHR histories, uncertainty surrounding diagnostic dates, unmodeled environmental and lifestyle exposures, and prediction testing limited to UKB restrict its current clinical interpretation.
Further mechanistic and prospective external validation using additional data sources is needed before the framework can support personalized risk profiling and precision medicine approaches.
Study links gut microbes to stronger cancer immunotherapy responses