The algorithm can evaluate the unconscious under general anesthesia in an accurate and reliable way.

Although anesthetics act on the brain, most anesthesiologists rely on heart rate, respiratory rate, and movement to speculate whether the surgical patient remains unconscious to the desired degree. In the new study, a research team based at MIT and Massachusetts General Hospital will use a simple artificial intelligence approach tailored to the type of anesthetic used to accurately and reliably unconscious patients based on brain activity. It shows that an algorithm to evaluate is obtained.

One of the things that comes to mind for anesthesiologists is, “Is there a conscious person lying in front of me who is unaware?” Ensuring patient awareness during surgery is fundamental to our activities. This is an important step forward. “

Emery N. Brown, Senior Author, Professor Edward Hood Taprin of Picower Institute for Learning and Memory and MIT’s Institute for Biomedical Engineering, MGH Anesthesiologist

Not only does Brown properly read the unconscious, but the anesthesiologist uses less medication than he would if he relies on direct, accurate, and unreliable indicators. It offers the potential to keep it at the desired level. This can improve postoperative outcomes in patients with delirium and the like.

“We may always have to be a little’outboard’,” said Brown, a professor at Harvard Medical School. “But can you do it with sufficient accuracy so that you don’t overdo it?”

For example, the algorithms used to drive infusion pumps help anesthesiologists precisely adjust drug delivery to optimize the patient’s condition and the dose they are receiving.

Artificial intelligence, real-world testing

Postdoctoral fellows John Abel and Marcus Badgley led the study to develop the technology to do so. PLOS ONE, We trained machine learning algorithms on a remarkable dataset collected by the lab in 2013. In that study, 10 healthy volunteers in their 20s were anesthetized with the commonly used drug propofol. As doses were systematically raised using computer-controlled delivery, volunteers were asked to respond to simple requests until they couldn’t. After that, I reduced the dose and regained consciousness, and I was able to react again. All the while, neural rhythms that reflect their brain activity were recorded on electroencephalogram (EEG) electrodes and provided a direct real-time link between the measured brain activity and the unconscious.

In a new task, Abel, Badgeley, and the team train a version of the AI ​​algorithm based on a variety of underlying statistical techniques in a snippet of over 33,000 2-second-long EEG records from seven volunteers. Did. In this way, the algorithm can “learn” the difference between EEG measurements that predict consciousness and unconsciousness under propofol. Next, the researchers tested the algorithm in three ways.

First, they checked whether the three most promising algorithms accurately predicted the unconscious when applied to EEG activities recorded from the other three volunteers in the 2013 study. They did.

The algorithm was then used to analyze EEGs recorded from 27 actual surgical patients who received propofol for general anesthesia. The algorithm was currently applied to data collected from “noisy” real-world surgical settings where rhythms were measured on various instruments, but the algorithm is more accurate than other studies have shown. Distinguished unconsciousness. The authors highlight one case in which the algorithm was able to detect a decrease in the patient’s level of consciousness minutes before the actual anesthesiologist did it. That is, it can be accurate and useful if used during surgery. Early alert.

As a third test, the team applied the algorithm to EEG records from 17 surgical patients anesthetized with sevoflurane. Sevoflurane, unlike propofol, is inhaled rather than infused, but functions in a similar manner by binding to the same GABA-A receptors on the same major types of brain cells. The team’s algorithm was run again with high accuracy, but with some reduction in accuracy. This suggests that the ability to classify the unconscious has been reliably inherited by another similarly functioning anesthetic.

The ability to predict the unconsciousness of different drugs with the same mechanism of action is important, the authors said. One of the major flaws in current EEG-based systems for monitoring awareness is that different categories of anesthetics work in very different ways and distinguish between drug classes even when they produce different EEG patterns. Don’t do it. They also do not adequately explain the known age differences in the brain’s response to anesthesia. These limits on their accuracy also limit their clinical use.

In a new study, the algorithm trained in the twenties was successfully applied to a cohort of surgical patients whose average age changed significantly over the years, but the authors found the algorithm for use in children and the elderly. I admit that I want to train clearly. You can also train new algorithms to specifically apply to other types of drugs with different mechanisms of action. All well-trained and tuned suites of algorithms may provide high accuracy in describing the patient’s age and the medication in use.

Abel said the team’s approach of assembling the problem of predicting consciousness through the EEG of a particular class of drug has made it much easier to implement and extend the machine learning approach.

“This is a proof of concept and shows that we can now look at older people and other types of drugs,” he said. “It’s easy to do this if you set it up the right way.”

The resulting algorithm is not computationally demanding. The authors stated that for a given two seconds of EEG data, the algorithm can make accurate predictions of consciousness in less than a tenth of a second running on a standard MacBook Pro computer.

According to Brown, the lab has already refined the algorithm based on the findings. He also said he would like to extend the test to hundreds of more cases to further examine their performance and determine if a wider distinction begins to emerge between the various underlying statistical models adopted by the team. I did.