ALADYNOULLI: A Bayesian Generative Model Integrates EHR and Genetic Data to Outperform Clinical Risk Scores in Disease Prediction
核心洞察
Researchers developed ALADYNOULLI, a Bayesian generative framework that integrates longitudinal electronic health records and polygenic risk scores to model disease risk trajectories across the lifespan.
The model demonstrated stronger discrimination of short- and long-term disease risk than established clinical scores including PCE, PREVENT, and the Gail model in UK Biobank evaluations.
ALADYNOULLI identified 151 genome-wide significant loci through signature-based GWAS and revealed distinct patient subgroups within broad diagnostic categories such as depression and breast cancer (搜索).
A novel Bayesian generative model called ALADYNOULLI has demonstrated the ability to outperform established clinical risk scores in predicting both short- and long-term disease risk by integrating longitudinal electronic health records (EHRs) with polygenic risk scores (PRS), according to a study published in Nature.
The framework, developed by researchers at Mass General Brigham (搜索) and collaborators, models age and EHR diagnoses alongside genetic data to reveal temporal patterns in disease risk and progression among diagnostic subgroups. Applied across three major biobanks—the UK Biobank (UKB), All of Us (AoU), and Mass General Brigham (MGB)—the model analyzed records from more than 683,000 participants covering 348 PheCode-defined disease phenotypes, with longitudinal records spanning up to 52 years.
"ALADYNOULLI captures shared disease patterns and achieves stronger risk discrimination than existing clinical scores in the evaluated comparisons," the authors wrote. "The model generated highly stable disease signatures across cohorts that were consistent with established clinical phenotypes in the examples directly examined."
Model Architecture and Disease Signatures
ALADYNOULLI represents the probability of disease occurrence as a mixture of probabilities across latent disease signatures. 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, with a median preservation probability of 80%.
The signatures reflected known biological characteristics of diseases. Individuals carrying familial hypercholesterolemia (搜索)-associated variants showed enrichment of the cardiovascular signature, while those with clonal hematopoiesis of indeterminate potential demonstrated stronger inflammatory signatures. Rare variant burdens in LDLR (搜索), TTN (搜索), and BRCA2 (搜索) aligned with their respective disease-associated patterns.
The model also learned disease progression patterns consistent with established medical knowledge, capturing clinically expected sequences in which hypercholesterolemia preceded myocardial infarction (搜索) and primary cancers preceded metastatic disease.
Superior Risk Prediction Performance
In head-to-head comparisons within the UK Biobank, ALADYNOULLI achieved higher areas under the curve (AUCs) for both short- and long-term disease prediction than established clinical risk scores, including the Pooled Cohort Equation (PCE) and PREVENT for atherosclerotic cardiovascular disease (搜索), and the Gail model for breast cancer (搜索). The model also performed favorably relative to Delphi-2M, an AI model for predicting individual diagnostic codes.
The framework's predictions complemented ICD code-level models by providing disease-level risk estimates rather than predictions for individual diagnostic codes. For inference on new patients, the vectorized implementation enabled rapid risk prediction—generating predictions for a single individual in approximately 0.05 seconds, or 8 minutes for 10,000 individuals.
Genetic Discovery and Patient Heterogeneity
Signature-based genome-wide association studies (GWAS) identified 151 genome-wide significant loci, including some cardiovascular associations that were not identified as lead loci in constituent cardiovascular GWAS. Heritability estimates for signature trajectories were calculated using linkage disequilibrium score regression (LDSC).
The model also revealed distinct patient subgroups within broader diagnostic categories. Descriptive clustering identified subgroups of patients with depression or breast cancer (搜索) who had greater loadings for inflammatory and metabolic disease signatures. For myocardial infarction (搜索), the cardiovascular signature loading rose more rapidly before the event in early-onset cases than in late-onset cases, potentially reflecting different biological mechanisms despite the same diagnosis.
Within the non-ischemic cardiovascular signature, modeled probabilities of atrial fibrillation (搜索) and heart failure (搜索) increased progressively after age 55. South Asian genetic ancestry was associated with greater cardiovascular signature loading, which peaked at approximately 50–60 years and remained elevated in older age.
Cross-Cohort Validation and Limitations
The model was trained and validated across three distinct biobanks, with the AoU and MGB cohorts serving as external validation datasets. The disease signatures demonstrated strong correspondence across biobanks, particularly for cardiovascular and malignancy signatures, "suggesting robust biological patterns that transcend population differences."
However, the authors noted several limitations. Incomplete EHR histories, uncertainty surrounding diagnostic dates, unmodeled environmental and lifestyle exposures, and prediction testing limited to UKB restrict the 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," the researchers concluded.
The study employed inverse probability weighting to address potential UKB participation bias while preserving core disease-signature relationships, and comprehensive washout analyses confirmed that predictions were not driven by diagnostic cascades immediately preceding events.