Harvard and Dana-Farber AI Tool Predicts Risk of 348 Diseases Using EHR and Genetic Data

Researchers from Harvard Medical School, Dana-Farber Cancer Institute, and Mass General Brigham have built an artificial intelligence model capable of predicting a patient's risk for 348 different diseases years before symptoms appear. The algorithm, detailed in a recent Nature publication, achieves this by combining a patient's electronic health record (EHR) history with their underlying genetic profile.[1][2]

The model, named ALADYNOULLI, departs from traditional diagnostic tools that evaluate one disease at a time. Instead, it uses a Bayesian generative framework to map out how interconnected conditions evolve over a patient's lifetime. This approach allows the system to continuously update its risk assessments as new medical data is added, effectively improving its predictive power as the patient ages.[1][2][5]

To build and validate the tool, the research team utilized data from over 683,000 individuals across three major databases: the UK Biobank, Mass General Brigham, and the All of Us research program. The dataset included up to 52 years of follow-up information, providing a deep longitudinal view of how diseases manifest and progress.[1][5]

By analyzing this massive dataset, ALADYNOULLI identified 21 "latent disease signatures." These are hidden biological patterns that capture underlying disease processes, such as metabolic-inflammatory pathways, which drive clusters of seemingly unrelated conditions.[1][5]

The integration of germline genetics is a critical component of the model's success. By incorporating polygenic risk scores alongside clinical data, the algorithm uncovered 151 genome-wide significant loci. Notably, this included cardiovascular associations that single-trait analyses had previously missed, demonstrating the value of a holistic, multi-disease approach.[1][5]

For clinical prediction, the tool has shown remarkable accuracy. It outperformed established risk scores across 28 conditions over both one-year and ten-year horizons. Specifically, it beat the Pooled Cohort Equation (PCE) and PREVENT models for cardiovascular risk, as well as the GAIL model for breast cancer risk.[1][2]

In practice, the AI captured clinically expected sequences—such as hypercholesterolemia preceding a myocardial infarction, or primary cancers preceding metastatic disease—but with added nuance. The model revealed that the cardiovascular signature rose more rapidly before a heart attack in early-onset cases compared to late-onset cases, suggesting different biological mechanisms are at play despite the same final diagnosis.[5]

At Dana-Farber Cancer Institute, researchers are already exploring how the tool can be applied to oncology. The team is using ALADYNOULLI to identify different patterns of cancer metastases, attempting to predict whether a tumor is likely to spread early, late, widely, or narrowly.[3][4]

This capability could help identify previously unrecognized subtypes of cancer. By understanding the biological reasons why patients progress along different metastatic trajectories, oncologists hope to identify more beneficial, targeted therapeutics.[3]

Despite these promising results, the evidence remains retrospective. The tool has currently only been tested on past patient data. Its real-world clinical utility, and whether these early warnings actually change doctor behavior or improve patient outcomes, has yet to be prospectively validated in active care settings.[2][5][6]

Furthermore, while the model uses an explicit likelihood formulation to account for selection bias through inverse probability weighting, AI tools trained on historical EHRs can still inherit systemic biases present in the underlying data. The researchers acknowledge that the signatures are concordant with established disease biology, but prospective clinical trials will be necessary to ensure equitable performance across diverse populations.[1][6]

The next phase of development involves expanding the model's signatures to increase the accuracy and biological grounding of its risk predictions. If successfully implemented in clinical practice, ALADYNOULLI could represent a significant step toward personalized, predictive medicine, allowing doctors to intervene years before pathology becomes clinically apparent.[2][3][6]