Trained with long-term observation data of 6 million Danes, the artificial intelligence model was able to predict with high accuracy important events in people’s lives up to the date of death. The accuracy of predictions can be further increased if observation data is also added to information about videos, correspondence, and social connections that accompany people’s lives. But first the ethical side of the issue needs to be resolved.

A joint project by researchers from the University of Copenhagen (Denmark) and Northeastern University in Boston (USA) showed that a transformer-type machine learning model can be used to predict events in people’s lives.

The transformer model was created to process strings such as natural language text. It differs from other models in terms of greater parallelization of tasks and does not require consistency in data analysis. It turned out that the model is very suitable for organizing data and predicting what will happen in a person’s life, and can even show the approximate time of death. The new model outperformed similar models previously created, at most, in terms of accuracy in predicting an individual’s behavior and time of death.

Article “Using the Sequence of Life Events to Predict Human Life” The description of the life2vec model created in the experiment based on the data of 6 million Danes was published in the magazine. Nature Computational Science. Also available for free on the website arXiv.org.

“We used the model to address a fundamental question: To what extent can we predict events in your future based on conditions and events in your past? From a scientific perspective, we are not interested in the prophecy itself, but in the nuances in the information that allow the model to provide such accurate answers.” – said Sune Lehmann, DTU professor and first author of the article.

The authors of the study used the order of events in people’s lives just like a sentence made up of words. For this very reason, the transformer model was used in the study created to analyze the texts. At the same time, the model works by taking into account known social laws and observations; Based on this, not only artificial intelligence, but also ordinary experts can draw conclusions about the further life path of a person depending on his place of residence, profession. , by social status, gender, habits and health card (visit to the doctor).

Data for training the Life2vec model were taken from labor market information and data from the National Patient Register (LPR) and Statistics Denmark. The dataset contains information on all 6 million Danes and includes information on income, wages, scholarships, job type, industries, welfare and more. The medical data set includes records of visits to healthcare providers or hospitals, diagnoses, patient type, and how sudden or urgent the request for medical care was. Data for the model is presented for the period from 2008 to 2020, but data for the restricted age group is from 2008 to 2016.

The authors of the study state that many ethical questions need to be answered for the full-scale use of such a model for social purposes. At the same time, they emphasize that common mechanisms for assessing the target audience in advertising allow for less information about people, and this is already in use. Therefore, there will be nothing wrong if the model can predict some negative events in the life of a particular person, which could have been prevented in one way or another. Meanwhile, the model predicts the date of death with an accuracy of four years.

According to the researchers, the next step will be to include other types of information in the model, such as text and images or information about our social connections. Using data in this way paves the way for a brand new interaction between social sciences and health sciences.