Researchers in Denmark have announced that they used powerful machine learning algorithms to predict certain aspects of human life, including the likelihood of an individual dying prematurely. Their study, published in the journal *Nature Computational Science*, explains how a machine learning algorithm model called "life2vec" predicted a person's life outcomes and behaviors when presented with highly specific data about them. Sunny Lehmann, the lead author of the study and a professor at the Technical University of Denmark, stated, "With this data, we can make any type of prediction." However, the researchers noted that it is a "preliminary research model" and cannot perform any "real-world tasks" in its current state.
Lehmann and his colleagues used data from a national registry in Denmark detailing a diverse group of 6 million people. The information included data from 2008 to 2016 concerning key life aspects such as education, health, income, and profession. The researchers adapted language processing techniques and created a vocabulary for life events so that the "life2vec" model could interpret sentences based on the data. Lehmann explained that "the algorithm learned from those data and became capable of making predictions about specific aspects of people's lives, including how they might think, feel, and behave, and even whether a person might die within the next few years."
To predict the early death of an individual, the team used data from January 1, 2008, to December 31, 2015, for a group of more than 2.3 million people aged between 35 and 65. Lehmann indicated that "the choice of this group was due to the difficulty of predicting deaths in this age category." The "life2vec" model used the data to infer the likelihood of a person surviving four years after 2016. Lehmann stated, "To test how well the life2vec model performs, we selected a group of 100,000 individuals, half of whom survived and the other half died." The researchers were aware of the individuals who died after 2016, unlike the algorithm. To test it, they had the algorithm make individual predictions about whether someone lived after 2016 or not. The results were impressive: the algorithm was correct 78% of the time.
The report emphasized that the "life2vec" model also outperformed other modern models and baselines by at least 11% by predicting mortality outcomes more accurately. Researchers found that males were at a higher risk of dying after 2016, as well as skilled workers such as engineers or people facing mental health issues like depression or anxiety, which also lead to early death. At the same time, holding a managerial position or having a high income often pushes people toward the "survival" column.
However, the research had several limitations, indicating that "the trials were not random, and researchers did not intend to allocate during the experiments and assess outcomes." The researchers only looked at data over an eight-year period, and there may be demographic social biases in the sample, even though everyone in Denmark appears in the national registry. They also noted that the study was conducted in a wealthy country with strong infrastructure and a robust healthcare system. It is unclear whether the results of "life2vec" could be applied in other countries like the United States, given the economic and social differences.
Lehmann acknowledged that the algorithm seems "ominous and crazy," but it has undergone significant work, especially by insurance companies. Dr. Arthur Caplan, head of the Division of Medical Ethics at NYU’s Grossman School of Medicine, agrees that insurance companies would be keen to outsmart consumers once models like "life2vec" become more commercial. He added, "This will make selling insurance harder. You cannot insure against risks if everyone knows exactly what the risks are."
However, Caplan, who did not participate in the new research, pointed out that the "life2vec" model does not predict the exact age at which a person will die or how. For instance, the algorithm cannot predict whether a person will be killed in a car accident. Caplan expects the emergence of more advanced predictive models within five years. He said, "We will have better institutions with larger databases that will provide suggestions on what to do to extend your life." Ultimately, Caplan argues that "using AI to predict when we might die removes the only aspect of our lives that keeps it interesting, which is mystery." He foresees advanced predictive models emerging in less than five years, explaining, "We will have better institutions with larger databases that will provide suggestions on what to do to extend your life." He further noted, "We are concerned about robots taking over the world and deciding they do not need us. But what we should be worried about is the manipulation of information by robots and their ability to predict our behavior, so we end up having a life so predictable that it robs us of some value from living."
