- An artificial intelligence (AI) model trained on data from the entire population of Denmark can predict chances of early death more accurately than existing models, including those used by the insurance industry.
- The Life2vec model was trained on data covering health, education, income, and occupation for 6 million people from 2008 to 2020.
- While there are potential applications in early prediction of social and health issues, researchers caution against misuse of the technology by big business.
The AI model, known as Life2vec, was developed by Sune Lehmann Jørgensen and his colleagues at the Technical University of Denmark. It was trained on a comprehensive dataset that included an array of personal details such as education, visits to doctors and hospitals, diagnoses, income, and occupation.
This dataset was converted into words to train a large language model, similar to that used in AI apps like ChatGPT. The researchers’ AI model predicts probabilities based on series of life events forming a personal history, just like predicting the next likely word in a sentence based on previous examples.
The model was trained using data between 2008 and 2016, with the last four years held back for testing. During testing, data on a group of people aged 35 to 65 were used, half of whom died between 2016 and 2020. The AI’s task was to predict who lived and who died. It demonstrated an 11% increase in accuracy compared to the existing AI models and actuarial life tables used in the insurance sector.
Moreover, Life2vec was also successful in predicting the results of a personality test in a subset of the population more accurately than AI models specifically trained for the task. Jørgensen believes that the model has enough data to aid in predictions encompassing a broad spectrum of health and social topics.
It is important to note that while this technology could be instrumental in predicting health issues early or helping governments reduce inequality, misuse remains a concern. Jørgensen emphasized the problematic potential of the model being utilized by insurance companies, which would disrupt the fundamental principle of insurance based on shared lack of knowledge. This prevention of misuse takes on added importance in light of information that similar technologies are already in use by large tech firms.
While the insurance industry has expressed interest in innovative predictive methodologies, the adoption of such technologies has been conservative and slow. This is because the consequences of errors could be significant in an industry where policies may remain in force for decades.
These findings are significant as they reveal the potential for AI to delve deep into the complexities of human life and draw accurate predictions about a range of possible outcomes, including early death.