Australian biologists have created an artificial intelligence (AI) system capable of predicting a person's lifespan with 69% accuracy from a single photograph of his organs, according to an article published in Scientific Reports.
Cybernetic "cuckoo"
In recent years, thanks to the development of mathematics and the growth of computing power of computers, scientists have the opportunity to create complex neural networks, artificial intelligence systems capable of performing non-trivial tasks and even “thinking” creatively, creating new examples of art and technology.
For example, in the last year alone, scientists have created AI capable of playing the "uncountable" ancient Chinese game of Go, searching the newspapers for the most important events in history, writing scripts for computer games, coloring photographs and videos "like Van Gogh" and drawing pictures. At the beginning of the year, scientists unveiled an AI system that can distinguish moles from skin cancer better than the most experienced dermatologists.
Oakden-Rainer and his colleagues took this idea further, creating a system of machine intelligence that can determine the duration of a person's life from photographs of his internal organs obtained with a computer tomograph.
This program is a so-called deep, or ultra-precise neural network - a multilayer structure of several tens or hundreds of simpler neural networks. Each of them does not process raw data, but analysis products obtained by the network located above, which makes it possible to simplify very complex problems and solve them using relatively modest computational resources.
These networks cannot solve problems immediately after they are created - like humans, they have to learn from their own mistakes for a long time before they start getting the right answers.
Promotional video:
The magic of artificial intelligence
For such training, Oakden-Rainer and his colleagues used a collection of several thousand chest and abdominal photographs taken with a tomography scanner during health observations of 40 patients. This set of images, according to scientists, was enough for their brainchild to be able to achieve the level of predictions that physicians usually demonstrate when trying to "by eye" determine the lifespan of their patients.
After making sure that the system they created correctly predicts the life expectancy from the photographs of the organs of already dead patients, the scientists checked how it would cope with the work in "combat" conditions. To do this, they recruited a group of eight young and elderly patients, illuminated their chest with a tomograph and observed their life over the next several years.
As it turned out, the program really coped well with the tasks assigned to it - it correctly predicted life expectancy for 69% of volunteers, correctly finding out which patients in the clinics would die in the next five years.
Since scientists do not know how such deep neural networks work "from the inside" and how they come to the conclusions, it remains unclear exactly what distinctive features the computer uses to predict the death of a person. At the same time, the relatively high accuracy of predictions for people suffering from obstructive pulmonary disease or heart failure, speaks in favor of the fact that such diseases most strongly influenced the "opinion" of AI.
Expanding the database and involving more volunteers in the experiments, scientists hope, will significantly improve the quality of predictions and make them more accurate for people who do not suffer from severe heart and lung diseases. Now, according to Oakden-Rainer, his team is "training" a new version of the neural network based on photographs of 12 thousand patients, which should significantly improve the accuracy of predictions.