Researchers have developed a system based on artificial intelligence that is able to predict death with 78% reliability.
Is it possible to predict everything until death? Danish and American researchers are sure of this. They developed an artificial intelligence model based on Transformers technology (an architecture used by ChatGPT) that is capable of predicting a person’s approximate date of death with great accuracy. If the basic principle seems disappointing at first glance, the intentions of the researchers are quite different. The latter aims to push the limits of understanding death by highlighting common characteristics that impair a person’s lifespan.
To create “Life2vec” (that’s its name), the researchers trained the model for several years on a large database containing detailed information on the lives of Danish residents over a 10-year period. In particular, the data include a labor market account containing information on employment, income, professional status, and the National Patient Registry, which contains all information about Danish residents’ contacts with the public health system, whether in a hospital or outpatient setting.
These two data sources were anonymized and merged at the individual level to create a “life sequence.” Each event in the resident’s life (medical visit, change in professional status, etc.) was included in this sequence with metadata such as the person’s age at that moment or a specific date. The model was then trained to understand life sequences and their sequence. Second, the model is specifically trained to predict mortality. The AI then observes the life sequence annotated with additional information (in binary form): whether the person died after 4 years or is still alive.
And the results are there. According to a full-scale test conducted in 2020 on 100,000 individuals aged 35 to 67, the Life2vec model is now able to predict a person’s date of death with 78.8% reliability, given their life history. According to the researchers, these results are better than conventional techniques for predicting death. Based on Life2vec data relayed by the University of Denmark, the researchers were able to identify common risk factors.
Surprisingly, older people (55 and over) are identified by the model as having a higher probability of mortality. Men also have a higher risk of death. Finally, socio-occupational category also plays an important role: working-class profiles are also more likely to die early. Conversely, profiles associated with a lower risk of death include women and people in management positions or people with higher incomes, as well as people without mental illness. For now, the researchers do not plan to make the model publicly available.
