Researchers have developed an artificial intelligence (AI) tool that improved doctors’ treatment of patients who developed sepsis, a deadly blood infection that can quickly shut down vital organs if not treated swiftly and correctly, according to a new report published in Nature Medicine.
Nearly 270,000 Americans die every year from sepsis—it’s the third leading cause of death—and “improving treatment even by a couple percentages [improvement] will save tens of thousands of lives each year [globally],” study author Anthony Gordon says.
Background: Sepsis occurs when the body has an infection somewhere that triggers a body-wide immune response that can lead to tissue damage, organ failure, amputations and death. Requiring quick response from a medical team, treatment usually includes antibiotics, intravenous fluids and sometimes a vasopressor drug to constrict blood vessels and raise the patient’s blood pressure. However, the timing and amount of fluids or drugs is tricky for doctors to know.
The researchers developed a machine-learning algorithm that looked retrospectively at data from 96,000 patients in U.S. hospital intensive care units. The tool, called AI Clinician, retrieved 48 variables for each patient, including demographics, vital signs, age and what fluids or vasopressors were already administered. It uses machine-learning called “reinforcement learning”—where the AI will compare a patient’s variables to those in its database to determine what the best amount of fluid and timing of vasopressor might be. Using a separate set of data, they then tested the algorithm against decisions made by doctors.
In comparing the results of the decisions AI Clinician would have made with the results from the doctors’ decisions, they found patients who had received the same treatment as suggested by AI Clinician had the lowest mortality rate, Imperial College London’s Gordon says. They also noted that where clinicians varied from the AI Clinician was, on average, to administer too much fluids and not enough vasopressor, but this depended on the individual patient.
