Researchers in the UW’s computer science and engineering department in collaboration with anesthesiologists at Harborview Medical Center, UW Medicine, and Seattle Children’s Hospital have developed a machine learning tool that can predict and help prevent hypoxaemia during surgery.
Hypoxaemia is a physiological condition described by low blood oxygenation that can occur during and after anesthesia use for surgery. Currently, anesthesiologists track real-time blood oxygenation, but there is a need to reliably predict and anticipate hypoxaemia so providers can take preventive actions and minimize patient harm.
To address this, Microsoft senior researcher Scott Lundberg and associate computer science professor Su-In Lee designed Prescience, a machine learning tool that predicts hypoxaemia risk during anaesthesia from multiple health-related factors. They trained their system on quantitative outputs from a hospital system that captures real-time data from patient monitors and anesthesia machines. Working alongside physicians, they identified additional static features (i.e. physical status, surgical procedures, and diagnoses) to use in their predictive-risk calculation, and how to design an interpretable clinical platform.
“Identifying those features is important; these are not linear associations,” Monica Vavilala, an anesthesiologist and the director of Harborview's Injury Prevention and Research Center, said. “To disentangle from patient level factors from procedural factors, it was really important.”
Vavilala provided a clinical assessment of the tool and its connection to anesthesiology workflow. As an anesthesiologist, she is responsible for making sure that a patient’s organs work well and that they do not experience pain during surgery. She accomplishes this by reading outputs from different monitors that track the body’s vitals, including pulse oximeters.
“This model that the paper developed will allow us to receive feedback from an intelligence system, based on an individual patient's history,” Vavilala said. “It will tell us what factors to address and in what order.”
Anesthesiologists have to weigh multiple factors simultaneously to identify appropriate care. The ability to anticipate, recognize early, and find patterns comes with practice and experience, and machine learning aims to capture this intuition.
For example, a patient’s hypoxaemia can be addressed during surgery, but after they leave the care of the surgical team, they can still experience respiratory arrest during recovery. By predicting the possibility of such an event with Prescience, physicians can place risky patients under appropriate care and administer different medications, in term saving more lives.
The outputs of this study showed that the predictive performance of Prescience was more accurate than that of anesthesiologists. However, the prediction accuracy of anesthesiologists improved when they evaluated their choices against Prescience’s explanations and risk prediction. Clinical acceptability is a critical challenge that such a technology needs to address.
“It questions the very foundational knowledge base you have developed. And after 25 years of experience, why would I let a machine tell me what to do?” Vavilala said. “The key here is to understand what our own limitations are and the limitations are of technology.”
Prescience is innovative because of how it presents such complex data that the algorithm evaluates; it explains why a prediction was made. Typically, such models comprise complexity and accuracy for more interpretable models.
“Sometimes data is presented to clinicians in the abstract,” Vavilala said. “[Prescience] would tell us that feature A has a greater risk than other features; it tells us what to target first.”
Reach reporter Vidhi Singh at email@example.com. Twitter: @vidhisvida
Like what you’re reading? Support high-quality student journalism by donating here.