Generative approach trained on UK and Danish data estimates disease risk over time and hints at a new era for prevention.
Imagine being able to map a person’s likely health trajectory twenty years into the future, not with certainty, but with probabilities calibrated across hundreds of possible conditions. That is the claim of a new generative AI model developed by EMBL, the German Cancer Research Centre (DKFZ) and the University of Copenhagen, which uses anonymized medical histories to estimate the timing and risk of more than 1,000 diseases. Trained on 400,000 participants from UK Biobank and tested on 1.9 million patients in Denmark, the model demonstrates that forecasting long-term health outcomes across two distinct healthcare systems is technically feasible [1].
The model, described in Nature, borrows from the playbook of large language models, treating medical diagnoses and lifestyle factors as sequential ‘tokens’ in a grammar of health. Just as an LLM learns sentence structure, this algorithm learns patterns of disease and comorbidity, forecasting what conditions might appear, in what order, and how quickly.
Longevity.Technology: Forecasting the arc of human disease with the same algorithmic finesse that predicts a sentence’s next word is both ingenious and fraught – ingenious because it allows trajectories of health to be simulated decades in advance, fraught because small missteps early on can cascade into outsized errors when stretched over twenty years. Reliability diminishes with distance; like weather forecasts, the nearer horizon is where the model shines. Yet even imperfect foresight is not without value – patterns of multimorbidity and disease clustering can be glimpsed, providing new opportunities to target prevention rather than treatment.
But the data foundations matter as much as the algorithms built upon them. UK Biobank participants skew older, whiter and healthier than the global average, while childhood and adolescent health is notably underrepresented; biases of this sort risk embedding systemic blind spots into the forecasts. For longevity science, the appeal is obvious – multimorbidity defines aging and the ability to model its emergence could sharpen both clinical and policy interventions – but without broader and more diverse inputs the promise of personalized foresight risks being unevenly distributed. The model is a proof of concept, not a panacea, and its most immediate utility may be in guiding population-level strategies rather than serving up individualized certainties.
How the model works
The AI is built on generative transformer architecture, customized with continuous age encodings to represent time between events. By capturing not just which conditions occur but also their order and spacing, it can generate plausible medical histories forward in time. Events include diagnoses, lifestyle factors such as smoking, and demographic features.
“Medical events often follow predictable patterns,” explained Tom Fitzgerald, Staff Scientist at EMBL’s European Bioinformatics Institute. “Our AI model learns those patterns and can forecast future health outcomes. It gives us a way to explore what might happen based on a person’s medical history and other key factors. Crucially, this is not a certainty, but an estimate of the potential risks.”
The model performs strongly on conditions with consistent progression such as certain cancers, heart attack and septicaemia, while proving less reliable for mental health disorders or pregnancy-related complications that are influenced by unpredictable life events [1].
Risk as probability
As with weather forecasts, the outputs are probabilities, not predictions. Forecasting horizons of a year are more accurate than those stretching two decades, but the estimates are well calibrated at population level. In the UK Biobank cohort, for instance, the annual risk of heart attack at ages 60–65 ranged from 4 in 10,000 for some men to 1 in 100 for others, depending on prior diagnoses and lifestyle, with women showing lower average risk but a similar spread [1].
“Our AI model is a proof of concept, showing that it’s possible for AI to learn many of our long-term health patterns and use this information to generate meaningful predictions,” said Ewan Birney, Interim Executive Director at EMBL. “By modelling how illnesses develop over time, we can start to explore when certain risks emerge and how best to plan early interventions. It’s a big step towards more personalised and preventive approaches to healthcare.”
From research to application
Like all models, Delphi-2M has limitations. Training on UK Biobank data means childhood and adolescent events are sparse, and the demographic skew raises questions about generalizability. The Danish validation helps, but further testing in diverse populations will be needed before clinical use. For now, researchers suggest the main benefits lie in studying how diseases develop, simulating outcomes when real data are scarce, and planning healthcare resources in the face of aging populations and chronic illness.
“This is the beginning of a new way to understand human health and disease progression,” said Moritz Gerstung, Head of the Division of AI in Oncology at DKFZ and former Group Leader at EMBL-EBI. “Generative models such as ours could one day help personalise care and anticipate healthcare needs at scale. By learning from large populations, these models offer a powerful lens into how diseases unfold, and could eventually support earlier, more tailored interventions.
Beyond the horizon
Generative AI in healthcare still faces hurdles – not least explainability, privacy safeguards and integration with clinical decision-making – but its trajectory is clear. By treating health histories as structured narratives rather than isolated datapoints, these models promise new ways to anticipate the transition from middle age to multimorbidity. If validated across more representative cohorts and combined with molecular and wearable data, such forecasting tools could become part of a preventive healthcare infrastructure that looks forward, not back.
