Nature Podcast · Springer Nature Limited

This AI tool predicts your risk of 1,000 diseases — by looking at your medical records

·36 min·3 clips

Moritz Gerstung reveals how Delphi2M predicts your lifetime risk of over 1,000 diseases using AI.

The Nature Podcast dedicates its first AI segment to a paper published in Nature describing Delphi2M, developed by Moritz Gerstung and colleagues at the German Cancer Research Center. The context Gerstung provides is that while hundreds of individual disease prediction algorithms have been developed over the past decades, each one targets a single condition: there is a tool for heart disease risk, a tool for diabetes risk, a tool for certain cancers, and so on. No tool has previously offered a comprehensive, simultaneous assessment of risk across the full range of conditions a person might face. Delphi2M addresses this gap by treating the problem analogously to how large language models are trained on text. Just as a language model learns to predict the next word given a sequence of prior words, Delphi2M learns to predict the next health event given a person's prior sequence of diagnoses, ages at diagnosis, and lifestyle factors. The training data comes from the UK Biobank, a cohort of approximately 500,000 participants who have contributed detailed health information over their lifetimes. The typical participant at the time of the study was between 60 and 70 years old, meaning the dataset includes full health trajectories from childhood diseases through the higher disease burden of later life. The system is input with a patient's health record and outputs predicted risks across over 1,000 disease categories simultaneously, with probability estimates that can inform preventive interventions. Reporter Benjamin Thompson asks Gerstung about the limitations, which include the fact that the training data is largely European and may not generalize well to other populations. The system also inherits the biases present in historical medical records, where certain conditions may be under- or over-diagnosed in specific demographic groups. The episode's second AI story covers how AI tools might facilitate academic dishonesty.

As heard by us

AI is framed first as a way to predict future illness, then as a test of whether machine use can encourage cheating.

Nature Podcast sets up two AI stories with different strengths. The first, on predicting future illness, has the cleaner arc: it takes time with the idea before widening to the medical stakes, which gives the reporting shape and keeps the science tied to everyday consequences.

Read the full review in PlayNext →

Why you'd press play

Press play if you want AI explained through future illness risk and a die-rolling cheating test.

Read the full recommendation in PlayNext →

Listen to the show on

Apple Podcasts Spotify

Play clips from this episode More from Nature Podcast