Nancy Young, MD, discusses predicting language outcomes after cochlear implant

Accurately predicting spoken language outcomes in deaf children who receive cochlear implants is critical to maximizing the benefits of this life-changing technology, according to Nancy Young, MD, professor at Northwestern University.

While cochlear implants are the first effective intervention to restore a human sense and have dramatically improved access to spoken language for many deaf children, outcomes remain highly variable when compared with children born with typical hearing. Some children develop strong listening and spoken language skills, while others progress more slowly and require intensive support. The ability to predict these outcomes early could transform both clinical care and research.

Young explained that reliable prediction would allow clinicians to identify, as early as possible, which children are likely to need more intensive or specialized intervention. This would support a more personalized approach to therapy, ensuring children receive the right “dose” and type of language intervention rather than a one-size-fits-all strategy.

In addition, prediction tools could help researchers design and test new behavioral therapies by identifying groups of children who are expected to struggle and evaluating which interventions work best for different brain profiles. Importantly, these predictions are grounded in brain structure, offering a biologically informed way to tailor care.

Traditional machine learning models, however, have important limitations. As highlighted in a 2024 Nature article, models that perform well in one population often lose accuracy when applied to different populations because they are sensitive to differences in training data.

This poses a major challenge in a global field such as pediatric hearing care, where children differ in language, culture, outcome measures, and imaging protocols. Retraining separate models for each country or language would be costly, time-consuming, and impractical.

To address these challenges, Young and colleagues explored deep transfer learning, an advanced form of deep learning designed to handle heterogeneous data. Their multicenter study included children from 3 continents, spanning different spoken languages, outcome measures, and magnetic resonance imaging (MRI) protocols. The models were trained using preoperative brain MRI data, hearing status before implantation, and post-implant language outcomes, with the goal of predicting whether a child would be a high or low language improver.

The results showed that deep transfer learning outperformed traditional machine learning approaches and remained accurate despite wide variation in data sources. One particular deep transfer learning model demonstrated strong performance across centers, suggesting the feasibility of a single, globally applicable prediction tool. Young emphasized that this “predict to prescribe” approach could accelerate translation into clinical practice and research.

Looking ahead, the team is recruiting additional centers worldwide to expand training datasets across more languages and settings. Beyond cochlear implants, this approach may eventually help identify children—both deaf and hearing—who are at risk for language delays even before speech emerges, opening the door to earlier and more effective intervention.

No relevant disclosures.

Reference

Wang Y, Yuan D, Dettman S, et al. Forecasting spoken language development in children with cochlear implants using preimplant magnetic resonance imaging. JAMA Otolaryngol Head Neck Surg. 2025. doi:10.1001/jamaoto.2025.4694