Anesthesia has been a part of medicine since the 1800s and has become an essential part of healthcare. Thanks to anesthesia, doctors can perform more invasive procedures that can save lives and improve our livelihood. Even something as ordinary as getting a root canal can be made better with the help of sedation.

Today, nearly 60,000 Americans receive anesthesia every day as they undergo a medical procedure. Anesthesiologists must manually calculate how much an individual should receive per dose to reach their desired state of unconsciousness. But what if there was a better, more accurate, and efficient way to determine an ideal dosage for an individual?

MIT researchers recently submitted a paper to the International Conference on Artificial Intelligence in Medicine (AIME) that sought to answer that question through a study investigating the potential for using deep reinforcement learning in patients who undergo anesthesia for a medical procedure.

Taking a Simulated Approach

In their paper, the MIT research team focused on the medicine Propofol, which is most commonly used in patients undergoing medical procedures that require sedation or general anesthesia.

So far, MIT researchers have not tested deep learning to control unconsciousness levels of patients under anesthesia in a clinical setting. Instead, the researchers developed and trained a deep neural network to control anesthesia dosing using reinforcement learning using a simulated environment.

The researchers could account for multiple patients with different features and characteristics by training their developed neural network on simulated patient data.

Potential Use Cases

The simulated events provided researchers with evidence that their advanced neural network is superior to previous methods of measuring the state of unconsciousness in patients. It outperformed the proportional-integral-derivative (PID) controller used previously to determine the ideal anesthetic dose for patients.

The researchers’ approach using a neural network has two distinct advantages. First, it allows for an enhanced ability to scale clinical variables in observation because of the relationship between the input variables and recommended dosage. Second, because deep neural networks allow researchers to create a model with continuous input data, the method produces more coherent control policies than previous policies, which were table-based.

Is This the Future of Anesthesia?

Basing anesthetic doses on more personalized data that comes directly from an individual’s brainwaves could provide greater accuracy for anesthesiologists who want patients to achieve a certain level of unconsciousness when undergoing a medical procedure. But is this study the beginning of the future of healthcare?

That answer will largely depend on what happens in controlled clinical studies if this research is allowed to move forward for testing on humans. The team of researchers hopes to move their trials to a clinical setting and test the neural network on humans soon. If they have the same results in a clinical setting as they do in a simulated environment, we might see their work being used out in the field in the future.