Reading patients' medical records is the basis of the new model'Delphi‑2M', a generative language model recently published in 'Nature'. Likewise how the ' workchatboxes' Like ChatGPT and Gemini—trained to complete texts probabilistically—Delphi transforms clinical data on diagnoses, age, sex, and lifestyle habits into tokens and learns temporal patterns to predict what diseases you might get—and when.
Trained on the UK Biobank database, Delphi aims to be a holistic model for simultaneously predicting more than a thousand diseases—with a more than remarkable average accuracy and surprising stability during external validation with Danish population data. But it doesn't just predict. It can also generate synthetic health trajectories. That is, simulate how a person evolves between the ages of 60 and 80, pointing out which previous diagnoses contribute most to each prediction. Its ultimate application would be to anticipate before diagnosis and open the doors to real primary prevention.
But the line between promise and hyperbole is fine. As cardiologist Eric Topol points out in his commentary, "'It's Time for Primary Prevention in Medicine'" Medicine has historically failed at truly preventing non-communicable diseases. It's true that we've improved the early detection of cancer or dementia, but preventing them altogether remains a milestone far from reality. Delphi-2M, according to Topol, represents a paradigm shift because it can individualize risk to individualize real risk. However, he also admits that these types of models have serious limitations: data selection bias, noise in coding and annotation, absence of critical contemporary layers such as genomics, and a total lack of prospective validation in real clinical settings as now of holistic prediction models. Implementing a system like this in clinical practice can lead to unforeseen side effects.Nature' They say it clearly: predictive potential cannot replace either the clinical or ethical interpretation of patients' contexts.
We will also have to consider how all this is regulated. What governance will these models have? What rights will patients have over their predictions? How will we prevent the perpetuation of inequalities in access to prevention?
Delphi-2M is a significant step forward. But it should only be a first step. To achieve true primary prevention, much more than predictions is needed. How to intervene in prevention must be thought out and managed, in a fair and safe manner. Here, technology must go hand in hand with regulation, medicine, and patient autonomy. Predicting diseases may help us see the future, but understanding how to act is another story.
