New AI model Life2vec may predict human lifespan: researchers



Researchers have unveiled a groundbreaking artificial intelligence (AI) tool named Life2vec, leveraging transformer models akin to those powering large language models like ChatGPT. Trained on a comprehensive dataset drawn from Denmark’s entire population, Life2vec utilizes sequences of life events encompassing health history, education, employment, and income to forecast various aspects, ranging from an individual’s personality to their lifespan.

The tool’s predictive accuracy, surpassing current state-of-the-art models, notably includes the ability to forecast individuals’ lifespans. However, the research team emphasizes that Life2vec should not be employed for real-time predictions on actual individuals. According to Tina Eliassi-Rad, a professor at Northeastern University in the U.S., the tool is a prediction model specifically tailored to a distinct dataset of a particular population and should serve as a foundation for future endeavors rather than a conclusive endpoint.

The development process of Life2vec involves collaboration with social scientists to infuse a human-centered approach into AI development. This collaboration aims to maintain focus on the human element amidst the vast dataset the tool has been trained on. Sune Lehmann, the study’s author published in Nature Computational Science, emphasizes that Life2vec offers a more comprehensive reflection of human existence compared to many other models.

At its core, Life2vec relies on an extensive dataset, utilizing long patterns of recurring life events to train the model. Taking inspiration from the transformer model’s approach used in training large language models, the researchers adapt it to represent a human life as a sequence of events.

The model learns from millions of life event sequences, creating vector representations in embedding spaces. These embedding spaces form the basis for the model’s predictions by categorizing and establishing connections between life events such as income, education, and health factors.

One notable prediction made by Life2vec is an individual’s probability of mortality. Visualizing the prediction space, researchers describe it as resembling a long cylinder, transitioning from a low to high probability of death. Notably, the model correctly identified instances where there was a high probability of death and the causes of death in cases with a low probability were often unpredictable, such as accidents.