A new study by researchers at MIT and Massachusetts General Hospital (MGH) suggests the day is approaching when advanced artificial intelligence systems could help anesthesiologists in the operating room.
In a special edition of artificial intelligence in medicine, the team of neuroscientists, engineers and doctors demonstrated a machine learning algorithm to continuously automate the dosing of the anesthetic drug propofol. Using a deep reinforcement learning application, in which the software’s neural networks simultaneously learned how its dosing choices maintain unconsciousness and how to critique the effectiveness of its own actions, the algorithm outperformed larger software. traditional in sophisticated physiology-based patient simulations. It also closely matched the performance of real anesthesiologists when showing what it would do to maintain unconsciousness given recorded data from nine real surgeries.
Algorithm advancements increase the ability of computers to maintain patient unconsciousness with no more medication than necessary, freeing anesthesiologists for all the other responsibilities they have in the operating room, including ensuring that patients stay still, feel no pain, remain physiologically stable, and receive enough oxygen, say co-lead authors Gabe Schamberg and Marcus Badgeley.
“You might think of our lens as analogous to an airplane’s autopilot, where the captain is always in the cockpit, attentive,” says Schamberg, a former MIT postdoc who is also the study’s corresponding author. . “Anesthesiologists must simultaneously monitor many aspects of a patient’s physiological state. So it makes sense to automate aspects of patient care that we understand well. »
Lead author Emery N. Brown, a neuroscientist at the Picower Institute for Learning and Memory and the Institute of Medical Engineering and Science at MIT and an anesthesiologist at MGH, says the algorithm’s potential to help optimizing drug dosing could improve patient care.
“Algorithms like this allow anesthesiologists to maintain more attentive, near-continuous vigilance on the patient during general anesthesia,” says Brown, Edward Hood Taplin Professor of Computational Neuroscience and Health Science and Technology at MIT.
Both actor and critic
The research team designed a machine learning approach that would not only learn how to dose propofol to keep the patient unconscious, but also how to do it in a way that would optimize the amount of drug delivered. They achieved this by equipping the software with two linked neural networks: an “actor” responsible for deciding how much drug to dose at any given time, and a “critic” whose job was to help the actor behave. in a way that maximizes “rewards” specified by the programmer. For example, the researchers experimented with training the algorithm using three different rewards: one that penalized only overdosing, one that challenged delivery of any dose, and one that imposed no penalty.
In any case, they trained the algorithm with simulations of patients who used advanced models of pharmacokinetics, or how quickly doses of propofol reach relevant regions of the brain after dosing, and pharmacodynamics, or how the drug actually alters consciousness when it reaches its destination. The patients’ levels of unconsciousness, meanwhile, were reflected in the brainwave measurement, as they can be in real operating rooms. By running hundreds of simulation cycles with a range of values for these conditions, the actor and critic could learn to perform their roles for a variety of patient types.
The most effective reward system proved to be that of the “dose penalty” in which the reviewer questioned each dose given by the actor, constantly berating the actor to continue to dose to the minimum necessary. to maintain unconsciousness. Without any dosing penalty, the system sometimes dosed too much, and with only an overdose penalty, it sometimes gave too little. The “dose penalty” model learned faster and produced fewer errors than other value models and the traditional standard software, a “proportional-integral-derivative” controller.
A competent advisor
After training and testing the algorithm with simulations, Schamberg and Badgeley put the “dose penalty” version through a more real test by feeding it patient consciousness data recorded from real cases in the operating room. . The tests demonstrated both the strengths and limitations of the algorithm.
During most tests, the algorithm’s dosing choices closely matched those of the treating anesthesiologists after unconsciousness was induced and before it was no longer needed. The algorithm, however, adjusted the dosage as frequently as every five seconds, whereas anesthesiologists (who all had lots of other things to do) typically only did so every 20 to 30 minutes, Badgeley notes.
As tests have shown, the algorithm is not optimized for inducing unconsciousness in the first place, the researchers acknowledge. The software also doesn’t know on its own when the surgery is complete, they add, but it’s a simple matter for the anesthesiologist to manage this process.
One of the biggest challenges any AI system is likely to continue to face, according to Schamberg, is whether the data it receives about patient unconsciousness is completely accurate. Another area of active research in Brown’s lab at MIT and MGH involves improving the interpretation of data sources, such as brainwave signals, to improve the quality of patient monitoring data under anesthesia.
Besides Schamberg, Badgeley and Brown, the other authors of the article are Benyamin Meschede-Krasa and Ohyoon Kwon.
The JPB Foundation and the National Institutes of Health funded the study.