AI helps predict and prevent chemotherapy-induced vomiting for kids with cancer

When four-year old Sam was undergoing treatment for Acute Lymphoblastic Leukemia (ALL), nausea was a constant. At one point, he was on three different medications — all of which made him feel sick.

“There were many times he would vomit out of nowhere,” says his mom, Tiffany. “And it’s not just puke ... it’s biohazardous waste. As a mom, your instinct is to jump in and help, but you have to be careful. It’s more than them feeling sick; it’s a safety issue.”

Now, a machine learning model developed in-house at The Hospital for Sick Children (SickKids) is helping families like Sam’s by predicting which children are at high risk of vomiting, enabling clinicians to intervene sooner.

“It could be the difference between a child having to receive nutrition through a feeding tube or being able to eat with their family or friends,” says Dr. Lee Dupuis, Clinical Pharmacist and Senior Associate Scientist, Child Health Evaluative Sciences at SickKids. “It’s a barrier towards living a normal life we can remove, one that makes everything just a little less isolating.”

Predicting risk before symptoms appear

The model is made possible because of Dr. Lillian Sung, Oncologist and Chief Clinical Data Scientist, and Dr. Adam Yan, Oncologist and Associate Chief Medical Information Officer (Academic Pediatrics), who have been spearheading Pediatric Real-world Evaluative Data sciences for Clinical Transformation, also known as PREDICT: an initiative at SickKids to develop, deploy, evaluate, and maintain machine learning models aimed at improving outcomes for patients and families.

Together, Drs. Sung, Yan and Dupuis zeroed in on chemotherapy-induced vomiting, building a model that can forecast a child’s vomiting risk within a 96-hour window. The model works by using data from the electronic health record (EHR), including information such as heart rate, blood pressure, and medications, to make a prediction that reflects each patient’s unique characteristics.

“As pharmacists, we always have an antiemetic plan in place for a patient to avoid or ease vomiting that their chemotherapy might bring,” says Dupuis. “The tricky part is that once a child has vomited, it’s more likely they will vomit again in the future ... meaning by the time it happens we may have already missed our window.”

The model is a clear reflection of Precision Child Health in action, a movement at SickKids to provide individualized care.

“Now instead of a plan that is based on the type of drug the patient is receiving, the model helps us make a plan that is based on the whole child in front of us,” adds Dupuis.

“Inpatient stays are always hard on kids, but the fact that they can catch vomiting ahead of time and prevent it is treatment-changing,” says Tiffany. “When Sam doesn’t feel nauseous, he wants to play, be active, socialize and go to the playrooms. It means I get to see my kid be a kid, not a sick child in a hospital bed.”

A recent silent trial demonstrated the model’s safety and efficacy and was published in BMC Cancer. A silent trial means the model runs in the background to make predictions, but the clinical team does not act on those predictions. With evidence of the model’s safety and effectiveness, the team is now using it to inform decision-making to improve vomiting control, alongside their clinical expertise.

PREDICT-ing a better future for patients and families

The vomiting prediction model is one of many that Sung and Yan are collaborating on as part of the PREDICT program, which aims to create robust, safe, bias-aware, and clinically trusted AI models that improve patient outcomes using EHR data.

That work depends on high-quality information, and the team spent two years converting the underlying EHR data into clinically relevant sources that could support machine learning.

“You need data to make predictions and to train models,” notes Sung. “That’s why we clean it, curate it and make sure that it’s right so we can start doing amazing things, like supporting this project in the oncology program. This data source, called SEDAR, is now available to SickKids staff who need access to reliable, high-quality data.”

The program has advanced to support three AI models nearing clinical deployment in oncology and cardiac care, and is part of a growing ecosystem at SickKids focused on developing, deploying and scaling responsible AI solutions.

That ecosystem includes programs like SickKids Artificial Intelligence (SKAI), which provides the enterprise-wide expertise and oversight needed to ensure AI tools are developed responsibly and in ways families and clinicians can trust.

This research was supported by the Leukemia Lymphoma Society of Canada, and the Cancer Research Society.