NTV: New neural network recognizes complex diagnoses more accurately than therapists
American developers presented artificial intelligence, which is much more accurate than ordinary doctors recognizes the most confusing diagnoses
American developers presented artificial intelligence, which is much more accurate than ordinary doctors recognizes the most confusing diagnoses. The program met clinical goals in 83% of cases, while GPs and GPs only 36%. The difference is more than twofold, and if the figures are confirmed in practice, medicine will face a major shift in the fight against diagnostic errors.
The program works on the principle of a medical consultation: several AI agents work inside it at once, each of which independently searches and compares available information. Their conclusions are then stitched together into a single answer, so the system simulates a collaborative discussion of a complex case by a group of specialists.
Test tasks were collected from medical journals: these are real clinical stories that are analyzed in life by specialized doctors. The neural network won in the vast majority of cases.
Skeptics recall details that spoil a beautiful picture. For the test, such complex cases were selected that in a real clinic they are considered by narrow specialists - and whole teams. But the specialized doctors themselves were not allowed to compete: they competed with therapists armed with reference books.
In other words, the comparison was not with the best representatives of the profession, but with the “middle league” on obviously “adult” tasks. Under other circumstances, the situation could have been different.
Even when adjusted for the conditions of the experiment, the result looks promising. The link between the doctor and the neural network – when the AI takes over the search for versions and rare coincidences, and the person checks and makes a decision – can significantly increase the accuracy of diagnosis. According to the WHO, misdiagnosis leads to millions of cases of mistreatment in the world every year, so even a partial reduction in error is a huge figure in the health system.
So far, we are talking about experimenting with journal materials, not working in real clinics. But the direction is set right: if further neural network will be honestly compared with narrow specialists, we will find out where it has a real ceiling.



