By Cami Rosso

Each day, more than 130 people die from opioid-related drug overdoses in the United States, according to estimates from the National Institutes of Health. The advancement of better practices for pain management to combat this crisis is one of the top priorities for the U.S. Department of Health and Human Services. 

In an effort to help clinicians provide better, more personalized pain management care, a research team led by Daniel Lopez-Martinez at the Harvard-Massachusetts Institute of Technology (MIT) Division of Health Sciences and Technology and MIT Media Lab, along with colleagues at MIT and Columbia University, announced last week the creation of an AI deep reinforcement learning algorithm for critical care pain management. They will be presenting their work on July 26, 2019, at the Engineering in Medicine and Biology (EMBC) Conference in Berlin.

A pain patient’s feedback to health care professionals about the intensity of pain experienced is important so that the dosage of pain medication can be adjusted during treatment. Excessive dosage may lead to addiction, and under-dosage may result in unmanageable pain and trigger chronic pain.  

Health care professionals face enormous challenges when it comes to effective pain-care management—especially in the intensive care unit (ICU) of a hospital where opioids are the preferred medication for treatment. ICU patients may not be necessarily able to communicate. This is often the case with children, intubated patients, the paralyzed, and those whose verbal capabilities are impacted by physical or cognitive damage.

Moreover, ICUs are frenetic environments, often punctuated by one life-threatening crisis after another. For a clinician to consider in real-time all the variables needed to provide optimal pain management is a daunting challenge. These variables may include the patient’s health history, physical traits such as gender and weight, any drug allergies, and expected duration of treatment.

The research team of Daniel Lopez-Martinez, Patrick Eschenfeldt, Sassan Ostvar, Myles Ingram, Chin Hur and Rosalind Picard took the approach of using AI reinforcement learning (RL) because, they wrote, “RL is able to infer optimal strategies from suboptimal training examples.” According to the researchers, this is the “first application of reinforcement learning for deducing optimal intervention regimens to manage pain.”

Deep reinforcement learning combines AI deep learning with reinforcement learning techniques. Reinforcement learning is inspired by behavioral science concepts, only in AI, an artificial (rather than biological) agent learns by interacting with its environment. It’s learning by trial-and-error motivated by reward or punishment, where over time, the artificial agent learns on its own what strategies result in achieving the greatest reward. Deep reinforcement learning has contributed to AI milestone achievements in chess, various Atari games, shogi (Japanese chess), and Go. In healthcare, deep reinforcement learning has been applied toward chemotherapy dosing for cancer patients in clinical trials, ICU heparin dosing, and the dosing of intravenous fluids and vasopressor for sepsis patients, among other purposes.

“We used data from more than 40,000 hospitalizations to understand what good and bad physician interventions with opioids may look like,” said Lopez-Martinez. “Based on this, our algorithm identified optimal actions, personalized for each patient. We believe that with additional work, we may be able to produce an AI system that can augment physicians' decision making so that pain medication is personalized for each patient. That will help prevent opioid addictions and combat the opioid epidemic.”

The researchers identified future possible ways to expand upon their proof-of-concept, such as including more data modalities, co-analgesic dosing regimes, and a broader range of therapeutic goals in the algorithm’s reward function, and expanding input data to include multiple centers.

“The goal would not be to replace physicians’ clinical judgments about treatment, but to aid clinical decision making with insights about optimal decisions and automatically guide therapy,” the researchers wrote.