Longevity Briefs: AI Could Predict Every Disease You’re Going To Get – And When

Longevity briefs provides a short summary of novel research in biology, medicine, or biotechnology that caught the attention of our researchers in Oxford, due to its potential to improve our health, wellbeing, and longevity.

The problem:

Prevention is better than cure, which means that healthcare decisions today often rely on understanding a person’s past and current health to predict their future. We have predictive models that are good at anticipating the occurrence of a single disease in isolation based on a reasonably narrow range of risk factors. In reality, however, people are usually in the process of developing (or have already developed) multiple diseases simultaneously. These diseases interact with each other and with the person’s environment in complex ways. Imagine if we could put all this information together and predict what diseases someone is going to get, when they are likely to get them and most importantly, what they can do about it.

It is in this kind of situation that machine learning could be of great use, by detecting intricate patterns within health data that hand crafted models would miss. The goal would be to produce a picture of an individual’s future health trajectory that might ordinarily require dozens of individual tests. In this study, researchers use modified GPT architecture – the same type of AI that powers ChatGPT – to model the progression and competing nature of human diseases. You may wonder how technology used to power a chatbot could be applied to human disease, but GPT is fundamentally a prediction tool. It predicts the next word in a sentence based on a vast dataset of interlinked words from human language, but can quite readily be used to predict the next health event in the same way, given enough health data.

The discovery:

Researchers created their new GPT-based model, named Delphi-2M, by training it on a large dataset from 0.4 million participants in the UK Biobank – an anonymised database containing detailed health information on British citizens. The model was then validated using both UK Biobank data that had not been used for training, and on data from 1.9 million Danish individuals to test its accuracy. It was given each person’s previous health history, demographic information (like age and sex) and lifestyle factors (such as smoking status) as its ‘prompt’. It was then made to predict the probability of different future health events, as well as the estimated time until said event would occur. This was compared to what actually happened to the validation populations to test the accuracy of the model.

The researchers found that Delphi-2M could predict the rates of over 1,000 diseases, with accuracy comparable to existing single-disease models, but in just a single ‘test’. However, the model still fared worse than some well established single-disease models. For example, HbA1c – a proxy measurement of the amount of glucose in the blood over an extended period of time – was still a better predictor of type II diabetes than Delphi-2M.

The implications:

By predicting the timing and progression of a wide range of diseases simultaneously, models like Delphi-2M could potentially help individuals and healthcare professionals alike to make more informed, personalised health and lifestyle decisions. They could be used to make predictions based on hypothetical data. Imagine, for example, being able to tell a specific person at what age they are likely to get cancer, heart disease, dementia or how long they are likely to live, and then showing them exactly how their health trajectory is expected to change with lifestyle interventions if they were to adopt them now. They could also be used to make predictions about health at the population level. For example, one could predict the expected rates of diseases for a given demographic in 30 years’ time and prepare for future healthcare needs.

It should be said that one important limitation of these machine learning approaches remains the availability and nature of the data used to train them. In this particular study, the UK Biobank training data included participants aged 40-70 at enrolment. This means that this particular predictive model might not work as well on the youngest and oldest of us. It was also observed that the model fared a little bit worse when predicting outcomes in the Danish population when compared to the British population, which highlights why the demographics in the training data matters.