By Deborah Borfitz
June 23, 2021 | Only about a quarter of anesthesiologists delivering anesthesia to patients during surgery are monitoring unconsciousness using electroencephalography (EEG), the recording of neural activity via electrodes on the scalp. Most anesthesiologists base the amount of anesthetic they administer on a combination of patient factors—vital signs, whether they move, and their medical history—as well as their own clinical experience.
This is due to training and the fact that current EEG-based indices force anesthesiologists to blindly accept the numbers produced by EEG-based spectral analyses indicating where patients’ brain activity falls on a scale of 100 (fully conscious) to zero (flatline), says Emery N. Brown, M.D., Ph.D. an anesthesiologist at Massachusetts General Hospital (MGH) and professor in The Picower Institute for Learning and Memory and the Institute for Medical Engineering and Science at Massachusetts Institute of Technology (MIT). The algorithms behind the numbers are proprietary.
“That is not good scientifically, it is not good for patients clinically, and it certainly impedes progress in the field,” says Brown, who has been working for years to get anesthesiologists out of this precarious situation with research tackling the 175-year neuroscience mystery of how anesthetics affect the brain. Three years ago, Brown and bioengineer Patrick L. Purdon, M.D., co-founded PASCALL Systems, Inc., a medical device startup developing novel technology spun off from their research labs at MGH and MIT.
The company’s tagline is, appropriately enough, Personalized Anesthesia State Control for All. Its ambitious, long-term goal is creation of a closed-loop EEG system where artificial intelligence is precisely and automatically controlling the amount of anesthesia drugs delivered to patients based on a known set of physiological parameters, Brown says.
MIT and MGH researchers recently published an article in PLOS ONE (DOI: 10.1371/journal.pone.0246165) detailing how they constructed elementary machine learning algorithms for real-time tracking of patients’ unconscious state during surgery that factors in the kind of anesthesia being used. The algorithms were shown to make highly accurate predictions of unconsciousness to accurately monitor anesthetic state.
Model Build
One of the main flaws of the current EEG-based systems for monitoring consciousness is that they do not distinguish among drug classes, which work in different ways and produce distinct EEG patterns, says Brown. They also do not adequately account for known age differences in brain response to anesthesia.
EEG-based spectral analysis was a concept introduced by Aspect Medical Systems, whose brain monitoring device was cleared by the U.S. Food and Drug Administration (FDA) for marketing in 1996. It was initially an engineering company looking for a biomedical use case for some signal processing algorithms it had developed, says Brown. “They had a hammer, and they were looking for a nail, and from that an answer sprung up.”
Presumably, all such systems are using regression analysis of data from patients in different states of anesthesia. Certainly, none of them state that anesthetics “change brain waves in a very systematic way,” he adds.
The approach of Brown and his colleagues was to apply what is known about a class of drug to infer how unconscious patients are based on their understanding of how anesthetics generate oscillations in the brain’s circuits. To illustrate this idea, the team trained machine learning algorithms on a data set gathered in 2013 from a group of 10 healthy volunteers, most in their 20s, who were given increasingly higher doses of one drug (propofol) while the neural rhythms reflecting their brain activity were recorded with EEG electrodes and they were asked to respond to a simple request until they could not anymore.
Different versions of the algorithms were trained on more than 33,000 two-second-long snippets of EEG recordings from seven of the original 10 volunteers so the models could learn to differentiate EEG readings predictive of consciousness and unconsciousness under propofol. Next, the researchers confirmed that the three most promising algorithms accurately predicted unconsciousness when applied to EEG activity recorded from the other three volunteers.
The algorithms were then used to analyze EEG recorded from 27 older patients who received propofol for general anesthesia in a “noisier” real-world surgical setting, where they distinguished unconsciousness with higher accuracy than shown in previous studies, Brown says.
Finally, the algorithms were applied to EEG recordings from 17 surgery patients who were anesthetized with sevoflurane, which is inhaled rather than infused but binds to the same GABA-A receptors on neurons in the brain as does propofol. The performance of the models was only modestly reduced, but still high, suggesting the algorithms could reliably classify unconsciousness broadly to other “GABAergic” drugs that account for about 92% of the ones used surgically, says Brown.
“There’s nothing magical about the machine learning,” he stresses. “It was just a way to relate the unconscious state of patients to variables that we thought were relevant.” The physiology—including what an EEG looks like for kids, young adults, older adults, and people who are very sick—had already been mapped out.
Patient Benefit
Well before a closed-loop EEG system automates anesthetic delivery, the new algorithms may be useful in the operating room to help anesthesiologists maintain unconsciousness at the desired level while using less drug than they might otherwise administer, Brown says. At present, anesthesiologists can look for a dramatic change in patients’ EEG pattern to be certain they are unconscious and may have some idea of how that pattern was generated, but the exact point where they transition from consciousness to unconsciousness or vice versa is “more of a challenge.”
The new algorithms could help anesthesiologists find the “minimal pattern” indicative of unconsciousness, he continues, so they can administer drugs to keep patients there without unnecessarily taking them any deeper. Anesthesia is overall very safe but also has side effects, and older adults are particularly vulnerable to postoperative brain dysfunction.
In addition to being better for patients, Brown adds, more precisely controlling anesthesia delivery is less costly. It might also make it possible for more surgeries to be done during the day.
Researchers are hopeful that the advent of a better and more transparent methodology for tracking patients’ brain activity will prompt anesthesiologists to more often use EEG recordings to monitor unconsciousness. A bonus for patients is that “accidentally” waking up in the middle of surgery—a fearful if rare complication of surgery—would be virtually impossible if anesthesiologists were opting to use the EEG machine, says Brown.
Over the near term, the new algorithm will offer anesthesiologists a clearer picture of the EEG analysis, but they will still need to do the interpretation, Brown adds. In another three or four years, the algorithm will probably be undergoing testing in an automatic control system with more intelligent, push-button functionality.
Studies of such closed-loop devices have taken place outside the U.S., but the markers specify a range and therefore are not precise, he says. The FDA has yet to approve any system for automatic dispensing of anesthetics largely because it is hard work to understand what the drugs are doing in the human body. “I think with the science we are doing such a device would be possible.”