New Machine Learning Tool Predicts a Child’s Personal Risk for Cisplatin-Induced Hearing Loss
The powerful chemotherapy drug cisplatin has been used since the late 1970s to treat a variety of cancers. It’s highly effective against solid tumors and is often a core element of treatment for children with brain and spinal cord tumors, neuroblastoma, and rhabdomyosarcoma.
Yet, cisplatin is well known to cause devastating side effects. In children, the most common side effect following therapy is debilitating hearing loss. Depending on the treatment plan, up to 80% of children treated with cisplatin end up with permanent hearing loss that affects their social lives, school performance, and future careers.
Now, an international team led by Etan Orgel, MD, at Children’s Hospital Los Angeles has developed a novel machine learning model that can predict an individual child’s risk of developing hearing loss from cisplatin treatment. Called PedsHEAR, the tool uses routine, readily available information to quickly predict this risk—with 95% confidence.
The team, which includes researchers from the Keck School of Medicine of USC and other institutions across the U.S. and Canada, is the first to develop and validate a novel machine learning model for this purpose.
Results were published in the Journal of Clinical Oncology, and the model is now available for public use.
A decades-long journey to personalize care
The study grew out of two decades of efforts to try to prevent cisplatin-induced hearing loss in children. Investigators from CHLA led the pivotal phase 3 Children’s Oncology Group trial of sodium thiosulfate (STS), and in 2022, the Food and Drug Administration approved STS as the first treatment to reduce the risk of hearing loss in children given cisplatin.
But patients’ treatment regimens are already highly complex, and some may not need STS to prevent hearing loss. For those who are not eligible for STS, it’s critical for clinicians to understand each patient’s risk and what options they have to protect that child’s hearing.
“We want to give families and providers the tools they need to understand their child’s risk and make an informed decision,” explains Dr. Orgel, who directs Quality and Patient Safety at CHLA’s Cancer and Blood Disease Institute. “This is the paradigm shift we're aiming for—speaking in certainties for each child versus speaking in generalities by regimen.”
This new predictive model is informed by a landmark study designed and led by Dr. Orgel in 2021. Researchers analyzed data from more than 1,400 cisplatin-treated patients across the United States and Canada to establish the first benchmarks for the prevalence of cisplatin-induced hearing loss in children and adolescents.
Researchers used the 1,400-person dataset as the foundation for their model, training it to analyze risk factors and probabilities and accurately predict a child’s risk level for hearing loss. The researchers also brought in two new, real-world data sets from the Children’s Oncology Group and a children’s hospital in Texas to validate the model in other populations. The now publicly available web model provides each patient with a percentage indicating the child’s individual probability of hearing loss.
Dr. Etan OrgelThis is the paradigm shift we're aiming for—speaking in certainties for each child versus speaking in generalities by regimen.
Machine learning approaches
Joshua Millstein, PhD, from the Keck School of Medicine of USC, led the creation and optimization of the highly complex machine learning model.
“We assessed a wide variety of modeling strategies to arrive at our final approach, which combines several machine learning methods, then applies a higher-level model—called an ensemble predictor—to integrate each model’s predictions into a single interpretable result,” he explains. “The main challenges of building the final model involved tuning it, which requires finding the model parameters that would optimize the tool’s performance.”
Recent advancements in ensemble predictor modeling helped the team overcome several challenges that have caused other models to fail in the past. “Ensuring that these models have enough patient data for pattern recognition can be exceedingly tricky when developing solutions for rare childhood cancers,” adds Dr. Millstein. “These new statistical techniques empowered us to deliver a more refined output, even with many differences between patients within our cohorts.”
Creating a new treatment planning standard
“My goal is for this to become a routine clinical tool,” says Dr. Orgel. “What's unique about this model is that it only uses routinely available data, so any doctor can use it from day one of diagnosis to plan treatment.
“Forewarned is forearmed going into chemotherapy,” he adds. “It’s so important to understand the options in front of you—and how to approach potential interventions with planned monitoring, such as frequency and compliance, with hearing testing.”
The research team’s next goal is to expand the model to young adults and adults up to 65 and to integrate genomics to make the model even more powerful.
“Ultimately, we aim to expand our approach to understand and predict risk for other common side effects of common chemotherapies,” says Dr. Orgel. “We want to equip all patients beginning their cancer journey with knowledge that supports meaningful discussions with their doctors on what to expect during and after treatment.”
Dr. Etan OrgelForewarned is forearmed going into chemotherapy. It’s so important to understand the options in front of you—and how to approach potential interventions with planned monitoring, such as frequency and compliance, with hearing testing.
