Imagine being able to tell an AI details about your life – your job, your eating habits, that time you broke an arm aged 4 – and have it predict the most likely health risks for people like you.
It sounds like pure sci-fi, but a new study demonstrates how this technology isn’t too far off. If you think about it, it’s not quite as scary as it sounds.
Scientists have long drawn parallels between how we live, eat, and sleep, and how that affects our risk of dying earlier or developing certain diseases. With an AI, that existing data just gets crunched faster and more comprehensively.
A team from Northeastern University, the Technical University of Denmark, and the University of Copenhagen in Denmark trained their AI engine on the data of six million Danish individuals to test the feasibility of this kind of prediction engine.
They found that the same training techniques that underpin Large Language Models (LLMs), powering bots such as ChatGPT, can also be applied to life events.
Instead of studying relationships between words and sentences, the AI figures out the relationships between everything that happens in our lives.
“The whole story of a human life, in a way, can also be thought of as a giant long sentence of the many things that can happen to a person,” says Sune Lehmann, a data scientist at the Technical University of Denmark.
“This model offers a much more comprehensive reflection of the world as it’s lived by human beings than many other models.”
The new model, dubbed life2vec, makes use of “embedding spaces” – where something in the real world is given a mathematical form that a computer can read – to draw out links between health factors, education backgrounds, income levels, and everything else that affects mortality rates.
When put to the test against known causes of death, the AI proved better than current methods at predicting how and when someone would die – although there are still plenty of events, like car accidents, that the model has no chance of being able to foresee.
The AI was also able to predict certain aspects of personality, such as extroversion.
Advanced AI tools like this show a lot of potential for spotting patterns that are too complex for humans to see, which means a better understanding of the relationship between how we live our lives and how healthy and well we are.
Despite these successes, the researchers are urging caution: the data only applies to Danish people for now and thus has embedded sociodemographic bias.
The team also wants to see concerns about privacy and personal data addressed before this is ever used in a real-world setting.
“Even though we’re using prediction to evaluate how good these models are, the tool shouldn’t be used for prediction on real people,” says computer scientist Tina Eliassi-Rad from Northeastern University.
“It is a prediction model based on a specific data set of a specific population.”
The research has been published in Nature Computational Science.
