A team of six undergraduate McGill students placed first in the International Artificial Intelligence Competition ProjectX, which ran from Sept. 1 to Jan. 31. Hosted by the University of Toronto, the annual competition challenges students to develop new models of machine learning with practical, real-world applications. Of the three categories open for submissions, including clinical practice, epidemiology, and genetics, the group placed first in clinical practice and received $25,000 in prize money.
The team worked to create DeepVent, a patient ventilator model that operates through reinforcement learning (RL)—a type of artificial intelligence (AI) that aims to recognize patterns in data to improve subsequent decision making.
Reinforcement learning is a small, yet rapidly developing subfield of AI and machine learning. When provided with a dataset the algorithm can begin to recognize patterns so that it can learn which decisions are more beneficial to reach a certain goal, similar to the way humans learn a new skill. Reinforcement learning is the same type of AI which now holds the title of best chess player in the world.
“It’s really impossible to [teach] all the different moves in chess,” Flemming Kondrup, U4 Science student and member of the winning team, said in an interview with The McGill Tribune. “So what you do instead is you play the agent against itself, and […] it learns a general pattern of how to play. That’s how humans learn as well, [because] they develop an intuition of what’s a good move and what’s a bad move.”
With this type of ‘intuitive’ analysis, RL can be used in medicine to learn from historical health data so that doctors can find the best, most personalized long-term treatment for individual patients.
“RL also looks into the future, and predicts how the treatment it’s going to give now is going to affect future treatment options,” Kondrup explained. “A really good chess player doesn’t just do a move that’s good now, but sometimes they might sacrifice a move now for a better move later. And the same thing applies to health care.”
The DeepVent team plans to use their model to regulate the dynamic ventilation needs of patients in hospital intensive care units. ICU doctors, who are often overworked and overwhelmed, may find it difficult to monitor the ventilation settings of multiple patients at once. DeepVent provides a solution to this through its personalized ventilation system that would adjust to changes in a patient’s breathing. Their results have shown a 59 per cent expected enhancement in overall treatment quality.
“As a disease evolves, you need to adjust the settings on the ventilator,” Kondrup said.“Doctors have to constantly monitor patients and adjust these settings, and that can be pretty challenging.”
The average hospital ventilator settings include functions to manage the fraction of oxygen inhaled from the surrounding environment (FIO2), the expiratory pressure when a patient breathes out (PEEP), and the volume of air inhaled (tidal volume), all of which can impact a patient’s respiration—and as a result, their overall health. DeepVent envisions a ventilator with a built-in feedback mechanism that self-adjusts the settings without doctors having to intervene.
Another challenge the team addressed is the extensive training and experience health-care professionals need to effectively manage ventilated patients.
“In order for a health-care practitioner to treat patients well, they need […] years of experience,” Kondrup said. “You can train AI on data of tens of thousands of patients [and] on what a real human being would need years to experience.”
As ventilators have been so central during the pandemic, it will be important to continue advancing the field of respiratory technology, both to help our intensive care units now, and into the post-pandemic future.
“The idea of DeepVent is that instead of just focussing on the present, it tries to promote long-term health and survival,” Kondrup said.
