Professional development

AI trained to treat blood poisoning and diagnose fractures

Vyacheslav Golovanov, Habr Effective treatment of patients requires a combination of training and experience.

This is one of the reasons why people are waiting with interest for the prospects of using AI in medicine: algorithms can be trained to use the experience of thousands of doctors, giving them more information than any person could digest.

At the end of October, there was some evidence that the software may have already come close to meeting these expectations. Two papers have been published describing excellent preliminary results of using AI for diagnosis and treatment. The papers indicate completely different tasks and approaches, which suggests that the range of situations in which AI can be useful is very wide.

Choice of treatment methods

One of the studies focused on sepsis (blood poisoning), which occurs when the immune system overreacts to infection. Sepsis is the third most common cause of death worldwide, and remains a problem even after hospitalization of the patient. There are methods of treating patients, but, judging by statistics, there are significant opportunities to improve the situation. Therefore, a small team of scientists from Britain and the USA decided to check whether the software could provide this improvement.

They used a reinforcement learning algorithm that is considered effective in situations with "rare reward signals." In other words, in such a large sample of the population, many other things will occur in the body, except sepsis, which will affect the results of any treatment, and therefore the signals of effective treatment will be weak and difficult to distinguish. This approach was developed to increase the chances of their recognition.

A large database was used for software training: more than 17,000 intensive care patients and 79,000 hospitalized patients from more than 125 clinics. The patient data contained 48 parameters of information, from vital signs and laboratory tests to demography. The algorithm used the data to determine the treatment that maximizes the patient's chance of survival for 90 days. The researchers called the resulting software an "AI clinician."

To assess the quality of the work of an AI clinician, a separate set of medical histories of patients was used. The algorithm was used to select a treatment method, after which the actual treatment of patients was compared with the proposed algorithm. In general, the PO recommended lower doses of injections and higher doses of vasoconstrictor drugs. People whose treatment coincided with such recommendations survived more often than other groups of patients.

Diagnostics

In the second work, the ability to detect problems requiring treatment, in particular, bone fractures, was evaluated. Often such problems are easy to see, but a small chip or a small crack is hard to notice even for a specialist. In most cases, the diagnosis falls on the shoulders not of a specialist, but of a doctor working in an ambulance. The new study does not seek to create an AI that replaces doctors, it only wants to help them.

The team asked 18 orthopedic surgeons to diagnose 135,000 images of potential wrist fractures, and then used this data to train an algorithm, a convolutional neural network with deep learning. The algorithm was used to mark the areas that doctors who are not specialists in orthopedics should pay attention to. In fact, he helped them concentrate on the areas where the fracture was most likely to occur.

In the past, such tests gave out too many diagnoses, and doctors recommended additional tests in harmless cases. But in this case, the accuracy of the diagnosis has increased, and false positives have decreased. Sensitivity (or the ability to detect fractures) rose from 81% to 92%, and accuracy (the ability to make a correct diagnosis) rose from 88% to 94%. In total, this means that the number of incorrect diagnoses of emergency doctors would be reduced by almost half.

In both studies, the software was not used in a context that fully reflects medical circumstances. Emergency doctors and doctors treating sepsis (and these may be the same people) will usually have a lot of additional reasons for excitement and distractions, so integrating AI into their work will be difficult. But the success of these attempts suggests that clinical trials of AI may begin earlier than previously thought, and after that we will really find out how much AI can help to make real diagnoses and prescribe treatment.

Portal "Eternal youth" http://vechnayamolodost.ru