Researchers have developed a single machine learning–based algorithm that uses longitudinal electronic health record (EHR) data and germline genetic information to simultaneously predict a patient’s risk of developing 348 distinct diseases over time, including several cancers. The findings were published in Nature.

“This tool offers a path toward improving the prediction of future diseases so doctors and patients can take action to try to prevent them,” co-senior author Alexander Gusev, PhD, of Dana-Farber Cancer Institute, Boston, commented in an institutional press release.

Study and Model Details

The investigators combined probabilistic modeling with machine learning to develop ALADYNOULLI, a Bayesian generative framework that uses longitudinal EHR diagnoses, age, and polygenic risk to identify latent disease “signatures” and model individual-specific health trajectories over time. Unlike approaches that treat diseases in isolation, the framework models disease occurrence as a weighted combination of signature-specific disease probabilities (ie, a mixture of probabilities rather than a probability of a mixture), allowing it to accommodate both simultaneous and chronic conditions.

“We have painstakingly curated these signatures, which is a big differentiator. In contrast to deep learning approaches, which are typically ‘black boxes,’ our curated signatures capture the underlying biology in an interpretable way,” stated co-senior author Giovanni Parmigiani, PhD, also of Dana-Farber Cancer Institute. “These signatures are then the drivers of the model’s ability to make predictions.”

ALADYNOULLI was applied to more than 683,000 patient records from the UK Biobank (n = 427,239), Mass General Brigham (n = 48,069), and All of Us (n = 208,263), with up to 52 years of follow-up. The analysis included 348 PheCodes mapped from International Classification of Diseases (ICD-10) codes.

The investigators used ALADYNOULLI for two complementary purposes, each involving a different analytical approach. For biomedical discovery, complete patient trajectories were leveraged retrospectively to characterize disease signatures, quantify genetic influences, and reveal patient heterogeneity within diagnostic categories; for clinical prediction, only information available up to each prediction time point was used to mirror real-world decision-making.

Key Findings

ALADYNOULLI recovered 21 replicable signatures that, according to the investigators, had high cross-cohort composition preservation (median = 80%) and revealed biologic subtypes within diagnostic categories (Cohen’s d up to 4.25; P ≤ 1 × 10⁻⁸ for 95% of comparisons).

Among patients with breast cancer, the gynecologic signature showed the strongest differentiation in one patient cluster (Cohen’s d = 4.25), whereas the pain/inflammatory/metabolic signature achieved near-complete separation in another cluster (Cohen’s d = 2.53).

“In clinic, two patients with the same diagnosis are not the same patient,” commented co-senior author Pradeep Natarajan, MD, MMSc, of Massachusetts General Hospital, Boston. “This model shows they often have different underlying signature profiles, which can translate into different progression patterns and different responses to the same treatment.”

The identified signatures were found to be concordant with established disease biology. For example, carriers of clonal hematopoiesis of indeterminate potential (CHIP) were enriched in the inflammation signature, whereas carriers of familial hypercholesterolemia were enriched in the cardiovascular signature. Rare variant burden in BRCA2, LDLR, and TTN also appeared to align with disease specificities and signatures.

When the investigators conducted a signature-based genome-wide association study, they identified 151 genome-wide significant loci, including cardiovascular associations not identified through single-trait analyses.

For disease prediction, ALADYNOULLI was found to outperform traditional clinical risk scores at 1- and 10-year horizons. In a breast cancer–specific comparison among women, the model achieved a 1-year area under the curve of 0.782 vs 0.549 with the Gail model. According to the investigators, its disease-level (PheCode) predictions complemented code-level approaches from foundation models such as Delphi-2M, which forecasts the risk of more than 1,000 diseases up to 20 years in advance.

Insights and Opportunities

The investigators noted limitations, including those related to biases inherent in EHR data (although the framework can address selection bias through inverse probability weighting), the exclusion of environmental and lifestyle factors, and the need for further functional validation of individual signatures. However, they concluded that “despite these limitations, ALADYNOULLI advances longitudinal health modelling for precision medicine by jointly capturing time-varying disease patterns and genetic predisposition, enabling patient subgrouping within diagnostic categories, and signature-level genetic discovery.”

As a next step, the team plans to train the model to identify different patterns of metastases in melanoma. They are also working to expand the model’s disease signatures to improve the accuracy and biologic grounding of its risk predictions, while exploring its potential use in clinical practice and clinical trial design.

“ALADYNOULLI not only can help predict different trajectories of metastases, but it also can give us a biologic understanding of why patients are progressing along these different trajectories,” Dr. Gusev said. “From a better understanding of the biology, we have the potential to identify more beneficial therapeutics.”

DISCLOSURES: The work was supported by a National Institutes of Health (NIH) National Heart, Lung, and Blood Institute (NHLBI) Mentored Clinical Scientist Research Career Development Awards (K08) grant, an American Heart Association Career Development Award, and a Burroughs Wellcome Fund award to Dr. Urbut; an NIH NHLBI grant to Dr. Nakao; a Wellcome Early-Career Award to Dr. Jiang; NIH grants to Dr. Natarajan; and a gift from the Dobson Family to Dr. Parmigiani. The study authors reported no conflicts of interest. For data and code availability, visit nature.com.

ASCO AI in Oncology is published by Conexiant under a license arrangement with the American Society of Clinical Oncology, Inc. (ASCO^®). The ideas and opinions expressed in ASCO AI in Oncology do not necessarily reflect those of Conexiant or ASCO. For more information, see Policies.