Artificial intelligence used to simulate written language can also be used to predict events in people's lives. Research projects from the Technical University of Denmark, University of Copenhagen, ITU, and Northeastern University in the United States show that if you use a large amount of data about people's lives and train "transformer models" (such as ChatGPT) that are used to process language, they can systematically organize data and even predict what events will happen in a person's life, and estimate their time of death.
In a recent article published in "Nature Computational Science" titled "Predicting Human Lives Using Life Event Sequences," researchers analyzed the health data of 6 million Danes and their attachment to the labor market in a model called life2vec.
The model, after an initial training phase where it learns patterns in the data, has been proven to outperform other advanced neural networks in accurately predicting outcomes such as personality and time of death.
"We use this model to address a fundamental question: to what extent can we predict your future events based on your past conditions and events? What excites us scientifically is not just the prediction itself, but the data that allows the model to provide such accurate answers," said Sune Lehmann, a professor at DTU and the first author of the article.
Predicting Time of Death
Life2vec's predictions are answers to general questions such as "death within four years." When researchers analyzed the model's answers, the results were consistent with existing findings in social science. For example, individuals with leadership positions or high incomes are more likely to survive, while being male, having professional skills, or having a diagnosis of mental illness are associated with a higher risk of death, all else being equal.
Life2vec encodes data into a large vector system, which is a mathematical structure for organizing different types of data. The model determines how to place data in terms of birth time, education, salary, housing, and health.
"What is exciting is to view human life as a long sequence of events, similar to sentences composed of a series of words in language. This is typically the task type that transformer models are used for in AI, but in our experiment, we use them to analyze what we call life sequences, which are events that occur in human life," said Sune Lehmann.
Raising Ethical Questions
The researchers behind the article point out that the life2vec model also raises some ethical issues, such as protecting sensitive data, privacy, and biases in the data. Before the model can be used to assess individual disease risks or other preventable life events, a deeper understanding of these risks is necessary.
"Similar technologies that predict life events and human behavior are already being used within technology companies, such as tracking our behavior on social networks, accurately profiling us, and using these profiles to predict and influence our behavior. This discussion needs to be part of a democratic dialogue so that we can consider where technology is leading us and whether this is the development we want," said Sune Lehmann.
The researchers explain that the next step will be to integrate other types of information, such as text and image data or information about our social connections. The use of this data opens up new interactions between social and health sciences.
