The study
This study introduces Delphi-2M, a model that examines the sequence of diagnoses and selected personal information to estimate the risk of subsequent events. It was trained on data from 402,799 UK Biobank participants and tested, without retraining, using health registries covering approximately 1.93 million people in Denmark.
The model considers more than a thousand diagnoses and can generate possible future trajectories. In the reported analyses, these simulations reproduced several population-level disease patterns over periods of up to twenty years. They are not certain forecasts of an individual person’s future: they describe possibilities and probabilities.
Why read it
The interesting idea is to study several diseases together, including their order in time, rather than treating each risk separately. Testing on data from another country helps assess whether a model can work beyond the population in which it learned. It also opens questions about how such models might inform healthcare planning.
Limits & open questions
This is a model-development and evaluation study, not a trial assigning patients to AI-guided care. It does not establish that using the model improves clinical decisions or health outcomes.
UK Biobank is not fully representative of the general population. The model can also learn distortions caused by missing diagnoses and differences in how records were collected. Associations between events do not establish cause and effect: predicting a risk does not automatically identify an intervention that would reduce it. Benefits, errors and consequences of using the model in clinical practice still require evaluation.