Last week brought a potentially significant step forward for early diagnosis in the form of a new AI model that can predict a patient’s next diagnosis across nearly 900 diseases using just real-world medical records.
Meet DT-Transformer. Researchers at Harvard, Brigham and Women’s, and the Broad Institute unveiled the GPT-style foundation model in a new arXiv preprint.
- DT-Transformer reads a patient’s medical history as a sequence and predicts which disease will show up next, and when it might come knocking.
The real headline is the training set. DT-Transformer was trained on 57.1M structured EHR entries from 1.7M patients across MGB’s 11 hospitals and 200 clinics (2000 to 2024) – the messy reality of U.S. clinical care rather than a polished research data set.
The numbers were impressive. A few standouts:
- DT-Transformer achieved a median AUC of 0.871 across 896 disease categories, with AUC over 0.5 for every condition.
- The model crushed an age- and sex-based baseline by +0.214 AUC (0.871 vs. 0.657), beating it on 96% of diseases.
- All of that was accomplished with a featherweight 2.2M parameter model that’s small enough to run just about anywhere (by comparison, Claude Fable 5 has about 6 trillion parameters).
The real test told a humbler story. When the team ran a true prospective test forecasting new diagnoses DT-Transformer had never seen, median AUC slipped to 0.713.
- That still beats the baseline on 80% of diseases, but that gap between the retrospective flex and the prospective reality is fairly significant.
- A 0.871 headline number and a 0.713 crystal ball aren’t the same product, and the second one is the one that patients would actually have to deal with.
One other highlight worth mentioning: including every repeated diagnosis worsened model performance, “drowning out the signal” rather than improving predictions – another reminder that more data isn’t always better. Better data is.
The Takeaway
Population-scale risk forecasting that runs on a model smaller than most phone apps is a real milestone, and training on routine records instead of a spotless biobank is exactly the kind of thing that could get models like DT-Transformer in front of actual patients – assuming the data holds up to peer-review.
