AI is good (maybe too good) to predict who will die prematurely

Medical researchers have unlocked a troubling ability in artificial intelligence (AI): predicting the premature death of a person.

Scientists have recently formed an AI system to evaluate a decade of general health data submitted by more than half a million people in the UK. Then they asked the AI ​​to predict whether people were at risk of dying prematurely – in other words, earlier than the average life expectancy – of a chronic illness, they reported in a report. new study.

The early death predictions that were made by the AI ​​algorithms were "significantly more accurate" than those provided by a model that did not use machine learning, the main author of The study, Dr. Stephen Weng, an assistant professor of epidemiology and data science at the University of Nottingham (UN) in the UK, said in a statement. [Can Machines Be Creative? Meet 9 AI ‘Artists’]

To assess the probability of premature mortality of subjects, researchers tested two types of AI: "deep learning", in which superimposed information processing networks help a computer to shoot examples; and "Random Forest", a simpler type of AI that combines several tree models to take into account the possible outcomes.

They then compared the findings of the AI ​​models to the results of a standard algorithm, called the Cox model.

Using these three models, scientists evaluated data in the British Biobank – an open access database of genetic, physical and health data – submitted by more than 500,000 people between 2006 and 2016. During this period, nearly 14,500 of the participants died, mainly from cancer, heart and respiratory diseases.

Different variables

All three models determined that factors such as age, sex, smoking history, and prior diagnosis of cancer were the key variables for assessing a person's likelihood of premature death. . But the researchers found that the models diverged from other key factors.

The Cox model relies heavily on ethnicity and physical activity, as opposed to machine learning models. In comparison, the random forest model put more emphasis on body fat percentage, waist circumference, amount of fruits and vegetables that people ate and skin tone, according to the report. ;study. For the deep learning model, the main factors were exposure to risks related to work, air pollution, alcohol consumption and the use of certain medications.

When all calculations were done, the in-depth learning algorithm provided the most accurate predictions, correctly identifying 76% of subjects who died during the study period. In comparison, the random forest model correctly predicted about 64% of premature deaths, whereas the Cox model only identified about 44%.

This is not the first time that experts are exploiting the predictive power of AI for health care. In 2017, a different team of researchers demonstrated that AI could learn to detect early signs of Alzheimer's disease. their algorithm evaluated brain tests to predict whether a person would be likely to develop Alzheimer's disease, and did so with a precision of about 84%, reported Live Science.

Another study showed that AI can predict the onset of autism in 6-month-old babies at high risk of developing the disease. Another study might detect signs of invasive diabetes through the analysis of retinal scanners; and another, also using data from retinal analyzes, predicted the likelihood that a patient would suffer a heart attack or stroke.

In the new study, scientists have demonstrated that machine learning – "with careful tuning" – can be used to successfully predict the evolution of mortality over time, said the co-author of the study, Joe Kai, professor of primary care at the United Nations.

Although the use of AI in this way may not be familiar to many health professionals, the presentation of the methods used in the study "could help in scientific audit and development future of this exciting field, "said Kai.

The results were published online today (March 27) in the journal PLOS ONE.

Originally published on Science live.