Predicting Disease Trajectories

Intriguing article by a team led by Moritz Gerstung (DKFZ German Cancer Research Center), Tom Fitzgerald and Ewan Birney (European Bioinformatics Institute | EMBL-EBI Hinxton, UK) reporting on the use of transformer model (to simplify, a modified version of GPT-2, an old model given the pace of change) trained on large sets of health records to predict disease progression, including co-morbidities, with a level of accuracy comparable to single-disease models.

Let's unpack: the model, Delphi-2M. predicts the rates of more than 1,000 diseases, conditional on each individual’s past disease history. It can create a cone of likely possible futures across many diseases. The model was trained on 0.4M UK Biobank participants and tested on unseen data from 1.9M individuals from Danish health records with no change in parameters, with only a minor degradation in predictive accuracy.

What is even more interesting (though knowing what might happen to you is definitely already exciting) is that, as the authors note, the model's "generative nature also enables sampling of synthetic future health trajectories, providing meaningful estimates of potential disease burden for up to 20 years, and enabling the training of AI models that have never seen actual data." For those of us interested in the diseases of aging, the applications are mind blowing, from both public health and precision medicine perspectives. Some of the figures, for example Figure 2 and other figures in the supplementary material showing multi-disease trajectories with age, leave me with a ravenous urge to experiment!