Delphi-2M, a new generative artificial intelligence (AI) model, estimates an individual’s risk for more than 1000 diseases up to 20 years in advance by analyzing electronic health records and lifestyle data. 

Developed by the European Molecular Biology Laboratory in Germany and the German Cancer Research Center in Heidelberg, the model uses a modified large language system to predict 1258 diseases based on medical history. These findings have been published in the journal Nature.

AI is already used in diagnostics and medical reporting; however, most predictive tools focus on a single condition. Delphi-2M enables multi-disease forecasting, addressing this gap and underscoring AI’s growing role in clinical decision support.

Model Development

Artem Shmatko of the German Cancer Research Center and colleagues trained Delphi-2M, a generative pretrained transformer model, on health data from 400,000 participants in the UK Biobank. 

It was subsequently assessed on records from 1.9 million individuals in the Danish National Patient Registry, which spans five decades of hospitalization. 

The model incorporates demographic and lifestyle factors such as age, sex, BMI, tobacco use, and alcohol consumption.

Accuracy by Disease

Delphi-2M achieved the highest accuracy for conditions with predictable courses, such as certain cancers and myocardial infarction. 

Performance was lower for psychiatric disorders, pregnancy complications, and rare diseases, with more variable disease trajectories. 

Overall, the predictive power matched or exceeded that of the single-disease models; however, the accuracy differed by condition.

Expert Views

“The model highlights the potential of generative AI in health research and possibly future clinical care,” said Robert Ranisch, PhD, a junior professor of medical ethics with a focus on digitization, Faculty of Health Sciences Brandenburg, University of Potsdam, Germany, regarding the results as discussed in the Science Media Center. 

He cautioned that uncertainty remains, noting that “bias and potential discrimination are central challenges for any AI model in medicine.”

Markus Herrmann, PhD, is head of AI ethics at the Institute for Medical and Data Ethics at Heidelberg University, Germany , described disease-risk prediction as a “necessary” application for this technology. “The potential benefits for patients and healthcare systems are enormous,” he said.

Carsten Marr, PhD, is director of the Institute of AI for Health at Helmholtz Zentrum München, and professor of AI in cellular therapy and hematology at Ludwig Maximilian University of Munich, noted broader applications: “What’s more interesting is identifying correlations between diseases that were previously unrecognized or detecting events that precede disease onset. For example, one study showed that Epstein-Barr virus infection increases the risk of multiple sclerosis by thirtyfold. These are the patterns we are looking for.”

Limitations and Ethics

According to Julian Varghese, MD, director of the Institute for Medical Data Science at Otto von Guericke University Magdeburg , Germany, Delphi-2M’s average area under the curve is 0.76. Data on short-term risks, such as whether a patient may develop lung cancer within one or 5 years, are lacking.

Ranisch emphasized that predictions must not be misinterpreted: “Patients must understand that these forecasts are not destiny. However, they can provide guidance for preventive or therapeutic decisions.” 

Herrmann added that informed consent is essential: “Ethically and legally, people have a right not to know — no one should live in fear of a possible disease.”

“The model is, first and foremost, just a tool,” Marr said. “How we want to use such a tool for our individual health data and how we weigh its potential against possible risks is a discussion that society must have. From my perspective, a well-regulated system like European healthcare is ideal for applying AI.”

This story was translated from Medscape’s German edition.