More than two decades ago, my father died at the dining table after what appeared to be a successful six rounds of chemotherapy for stage 4 cancer (non-small cell lung carcinoma). It wasn't until 2011 that there were enough studies to understand that cisplatin and carboplatin, the standard treatment for cancer at the time, do indeed cause an increased risk for patients to experience stroke. What likely killed him in such a violent manner boils down to a lack of data & compute.
Over time, as better data and more computation capability have improved, we have demonstrated that his life and violent death could have been prevented if we had had the data to compute the right treatment for him - not for the average patient, but specifically for him. For decades, that vision was a pipe dream.
Today, I'm pleased to announce that we have started training on the greatest clinical data set the world has ever assembled in any one place. Using de-identified longitudinal medical records spanning nearly two decades for more than 40 million patients, we extracted the signal on cleansed, de-duplicated and encoded records that includes 9.1 billion observations. Using this we are training a world model of human health that will allow new discoveries and personalized medicine at a scale infeasible up to this moment. Importantly, the data set isn’t static - it is updated with new data from those patients as they interact with the health systems they use.
This is not our first time doing this. The team behind this effort has built Foresight in the UK as well as a deep background in data privacy, clinical medicine, and the healthcare records infrastructure in the United States. We spent the summer proving that we can extract more signal out of clinical data than the state-of-the-art.
Our vision for the world model is to approach this in two steps:
A large health model where instead of training a model on English language, we are teaching the model to predict the next clinical event in medical terminology.
The analogy I like to draw is that using LLMs is like using the smartest English student that has read every medical textbook and every medical paper and is now your doctor - but has never practiced medicine.
By contrast, a large health model is a doctor that is the most experienced doctor the world has ever seen - someone who has practiced 60 million cases already, and is trained in the medical literature, and is now using that body of experience for the patient in front of them.A world model of human health that is able to simulate the patient’s future trajectory. It is not a chatbot that quotes guidelines, and it is not a one-shot risk score. It is an internal simulator: a learned model of how health states change over time, including what happens if you intervene. A personal health world model would keep a compact state of you - labs, imaging, wearables, sleep, glucose, medications, microbiome, behavior - and generate plausible futures.
Today’s typical health models |
A health world model |
Generates text or classifies an image or predicts a label |
Generates a clinical trajectory |
Answers “what is this?” |
Answers “what happens if we do X?” |
Narrow - One task, one dataset |
Shared state used for forecast, risk, simulation, planning |
Language models reciting knowledge |
Dynamics that stay consistent across many steps |
Observes care |
Can be used as an environment to practice care |
We’re calling this Vitruvia - Leonardo da Vinci applied the greatest quantitative techniques of the fifteenth century - compass, caliper, and the patient measuring of living bodies - to inscribe man as a microcosm of the cosmos; now we extend that same work into a world model for human health, a living simulation that seeks to hold every system, every variation, and every possible future of the body within its computational cosmos.
This is the most exciting and meaningful mission I've ever been on in my life. If you would like to collaborate, please reach out.