Bayesian model combining genetic and health records outperforms clinical risk scores

A team of researchers has built a new kind of disease prediction model that works differently from the clinical risk scores doctors have relied on for years. Rather than asking whether you're likely to have a heart attack or develop breast cancer in isolation, this model—called ALADYNOULLI—looks at the interconnected web of diseases that tend to cluster together in human bodies, and it does so by weaving together your electronic health records, your age, and your genetic predispositions. When tested against established clinical benchmarks using data from more than 683,000 people across three major biobanks, the model outperformed those older scores at predicting both near-term and decade-long disease risk.

The researchers published their work in Nature after applying the framework to records spanning up to 52 years of medical history. The model identified 21 distinct disease signatures—patterns of conditions that tend to occur together—by analyzing 348 different disease phenotypes drawn from electronic health records. What makes this approach novel is that it doesn't treat each diagnosis as an isolated event. Instead, it recognizes that a person with high cholesterol, for instance, exists within a broader cardiovascular signature that also encompasses atrial fibrillation, heart failure, and myocardial infarction, each with its own age-dependent probability of occurring. The model learned these patterns from the data itself, discovering that high cholesterol typically precedes heart attacks, and that primary cancers precede metastatic disease—sequences that align with what clinicians already know but that a purely statistical model had to infer from millions of patient records.

One of the model's strengths lies in its ability to reveal hidden variation within diagnostic categories that appear uniform in conventional medicine. When researchers looked at patients with depression or breast cancer, for example, they found subgroups that carried much higher loadings for inflammatory and metabolic disease signatures than others with the same diagnosis. This suggests that two patients with identical diagnoses may actually be traveling different biological pathways, a distinction that could matter for treatment decisions. The model also uncovered genetic associations that single-disease analyses had missed. In one example, rare genetic variants in the LDLR gene—known to cause familial hypercholesterolaemia—showed up enriched in the cardiovascular signature, while variants in BRCA2 aligned with cancer-related patterns. The genome-wide association studies performed on the disease signatures themselves identified 151 genome-wide significant loci, some of which had not appeared as lead variants in conventional cardiovascular studies.

The model's predictions also tracked with biological reality in ways that suggest it has learned something true about how disease risk unfolds across a lifetime. Within the non-ischaemic cardiovascular signature, the probability of atrial fibrillation and heart failure rose sharply after age 55. Within the malignancy signature, the risk of metastatic disease spiked between ages 60 and 75. South Asian genetic ancestry was associated with greater cardiovascular signature loading that peaked around 50 to 60 years and remained elevated thereafter. These patterns held up consistently across all three datasets, suggesting the model had identified something reproducible rather than an artifact of a single population.

When the researchers directly compared ALADYNOULLI to established clinical risk scores—the Pooled Cohort Equation and PREVENT for cardiovascular disease, the Gail model for breast cancer—the new model achieved higher areas under the curve for both one-year and ten-year predictions in the UK Biobank cohort. The model also outperformed Delphi-2M, an AI system designed to predict individual diagnostic codes. This matters because better discrimination between high-risk and low-risk individuals could help clinicians identify who needs closer preventive monitoring and earlier intervention.

Yet the researchers are cautious about immediate clinical deployment. The model was trained and validated primarily on UK Biobank data, and it does not account for environmental exposures or lifestyle factors that shape disease risk in the real world. Electronic health records are often incomplete, with diagnostic dates that may be uncertain, and some conditions are simply underrepresented in the data. Before ALADYNOULLI can guide personalized risk profiling in clinical practice, it will need prospective validation—testing on new patients followed forward in time—and external validation using additional data sources beyond the three biobanks already studied. The framework shows promise as a tool for precision medicine, but the path from research finding to clinical tool remains open.