Johns Hopkins All Children’s Researchers Develop Predictive, ‘Machine Learning’ Models to Make Pediatric Anesthesia Even Safer
Over the years, the administration of anesthesia to pediatric surgical patients has risen to a very safe level. However, thanks to recent research carried out by Johns Hopkins All Children’s Hospital investigators and their European colleagues, the parents of a child who is going to have a surgical procedure will likely have even fewer worries about the safety of anesthesia.
“While the safety level for pediatric patients receiving anesthesia for surgical procedures has risen to very high levels, there remain small chances of serious complications, even for patients who have already been classified as being at low risk for adverse anesthesia side effects,” explains Geoffrey M. Gray, Ph.D., an assistant professor of anesthesiology and critical care medicine at the Johns Hopkins University School of Medicine based on the St. Petersburg, Florida, campus.
Accordingly, Gray, who is also a member of Johns Hopkins All Children’s Center for Pediatric Data Science and Analytic Methodology, and his colleagues, employed “machine learning,” a sub-type of artificial intelligence (AI), to develop two computerized models that can “fine tune” the screening of individual pediatric patients who — using other screening methods — had already been classified as being at low risk for adverse effects from anesthesia by another scoring system.
The new screening models developed using machine learning can offer the pediatric surgical team a more individualized and comprehensive view of their patients’ low risk at two points — when the surgery is “booked,” and repeated on the day of surgery.
What is ‘Machine Learning’ and How Can it Improve on Determinations of ‘Low Risk?’
“Machine learning is being used increasingly for predicting outcomes in medicine,” says study co-author Luis M. Ahumada, Ph.D., director of Predictive Analytics and the Center for Pediatric Data Science and Analytic Methodology (CPDSAM), who helped design the study and analyzed the data. “Machine learning uses algorithms — computerized procedures used for solving problems or performing computations — which can be used to discover and help visualize patterns as well as make predictions.”
Ahumada adds that the algorithms that are the basis of machine learning analyze and interpret data that can be employed in models that can be used as clinical decision-support tools and that the computerized tool can receive, learn and improve its knowledge without additional programming.
Ahumada, who is also a St. Petersburg-based assistant professor at Johns Hopkins University School of Medicine, adds that AI and machine learning do not represent an effort to replace your doctor with a computer; however, the physician and the model working together can be very powerful.
“Using predictive analytics, we can find solutions for developing clinical decision-making support systems, to aiding in disease management, in tracking hospital readmissions, or gathering information regarding the causes of adverse events and subsequently preventing them,” Ahumada says.
Using the APRICOT Dataset
A paper describing their research, titled: “A machine learning approach for decision support and risk stratification of pediatric perioperative patients based on the APRICOT dataset,” was recently published in the journal Pediatric Anesthesia.
According to Gray, the “Anesthesia PRactice In Children Observational Trial study”(APRICOT) dataset, was derived from a prospective, observational study conducted by 261 European institutions and included over 30,000 pediatric patients up to age 16 who were undergoing elective or emergency surgery or diagnostic procedures.
The researchers developed two models to improve upon the one provided by the American Society of Anesthesiologists Physical Status (ASA-PS) Classification System, used globally to determine risk for adverse events. The ASA-PS system is not based on individual pediatric patient features and, according to the researchers, often offers inconsistent determinations.
According to the study authors, the ASA-PS offers a “population” perspective on risk while the machine learning models offer an “individual” perspective, one where individual pediatric patient health and demographic features are programmed into the machine learning process, offering an improved, “personalized medicine” approach.
Clinicians may use the ASA-PS to classify patient risk, but it was not designed for that purpose, says Gray.
“The authors of the APRICOT data set did not aim to build a risk stratification tool,” explains Gray. “However, our goal was to generate risk prediction models by ‘leveraging’ the health-related features in the APRICOT dataset with machine learning techniques to develop models for fine tuning, which pediatric patients who had already been classified as being at low risk for adverse events related to anesthesia.”
Including Important Features Data
Two sets of “features,” both health-related and social, were selected with input from subject matter experts to ensure the “clinical utility” of the models.
“A model’s output of ‘low risk’ for an individual patient could provide additional reassurance for the surgical team,” explains Gray.
The two models use 30 health and social features about an individual pediatric patient who will be receiving anesthesia. The health-related features included “asthma,” “allergy,” “fever,” “medications,” among many others.
They also added whether the upcoming procedure was elective, urgent or emergency.
Social and demographic features included age, sex, weight and whether any family members smoked. Other features programmed included information regarding what kind of anesthesia was to be used, the experience of those providing anesthesia, how the anesthesia was to be administered and the time of day at which the surgical procedure was scheduled.
“In the booking model, the most important predictors were procedure time, family history of smoking, patient age, anesthesia type and snoring,” explains Gray. “In the day of surgery model, the most important features were urgency, parental presence, procedure time, patient weight and anesthesia type.”
The value in having two models can allow for modification — such as regarding the type of anesthesia used, says Gray, adding that the models were “trained” to use only the data that would be available at the two time points — the time of surgical booking and on the day of the surgical procedure.
Future Research Using Machine Learning to Determine Anesthesia Low Risk
“Our research demonstrated that prediction of patients at low risk for perioperative adverse events can be made on an individual level rather than a population level using machine learning,” reiterates Gray. “By using machine learning predictions generated based on individual patient factors, rather than using population-based estimates generated by using the more general ASA-PS, we can improve upon the potentially high variability and misleading risk estimates that have been associated with ASA-PS scaling.”
According to Gray, although their models were developed for “fine-tuning” low-risk classifications for pediatric surgical patients receiving anesthesia, their future work could focus on developing risk-based stratifications for all patients, including the elderly, for whom anesthesia may also be accompanied by certain risks.
The study was funded internally by the Department of Anesthesia and Pain Medicine at Johns Hopkins All Children’s Hospital and the APRICOT dataset was used with the approval of the APRICOT steering committee and European Society of Anesthesiology and Intensive Care (EASIC).