More accurate than the insurance industry’s tools, a model trained on the health, employment, and financial information of six million people can forecast death.
The probability of a person dying may be predicted more precisely by an artificial intelligence educated on personal data about the whole population of Denmark than by any current model, even those employed by the insurance sector. The technology’s developers believe that it might also be beneficial in the early detection of social and health issues, but it has to be kept out of the hands of large corporations.
Technical University of Denmark researcher Sune Lehmann Jørgensen and colleagues examined a comprehensive Danish dataset including 6 million people’s income, occupation, visits to physicians and hospitals, education, and any subsequent diagnosis between 2008 and 2020.
The same technology that powers ChatGPT and other AI apps was transformed from this information into words that could be used to train a big language model. These models function by analyzing a list of words and using a large sample size to predict the word that is statistically most likely to appear next. Similarly, the Life2vec model developed by the researchers may predict future occurrences based on a set of life events that make up an individual’s past.
Life2vec was trained in trials using all data save the latest four years, which were reserved for testing. Using data from a group of individuals between the ages of 35 and 65, half of whom passed away between 2016 and 2020, the researchers asked Life2vec to forecast which of the individuals would survive. Compared to the actuarial life tables used in the financial sector to price life insurance policies, it was 11% more accurate than any other AI model currently in use.
Additionally, the model outperformed AI models developed expressly for the task in predicting a portion of the population’s answers to a personality test.
According to Jørgensen, the model has ingested so much data that it should be able to provide insight into a variety of social and health-related issues. This means that governments may utilize technology to lessen inequality or to anticipate health concerns and catch them early. However, he emphasizes that businesses might potentially utilize it negatively.
“It is obvious that an insurance firm should not utilize our concept, as the entire purpose of insurance is to share the burden of not knowing who will be unlucky enough to be involved in an accident, die, or misplace their backpack,” explains Jørgensen.
However, he points out that similar technologies are currently available. “They’re probably already being used against us by large tech companies that have a ton of data about us and use it to predict things about us.”
Insurance companies are undoubtedly interested in novel predictive techniques, according to Matthew Edwards of the Institute and Faculty of Actuaries, a professional organization in the UK. However, he notes that most decisions are made by generalized linear models, a primitive form of AI compared to this research.
“Taking what data they have and trying to anticipate life expectancy from that is what insurance firms have been doing for many, many tens or hundreds of years,” adds Edwards. However, we purposefully take a cautious approach when implementing new procedures because the last thing you want to do when drafting a policy that may be in effect for the next 20 or 30 years is to make a serious error. Change is possible, but it happens slowly as no one wants to make a mistake.