Forecasting Spoken Language Development in Children With Cochlear Implants Using Preimplant Magnetic Resonance Imaging

Importance Cochlear implants substantially improve spoken language in children with severe to profound sensorineural hearing loss, yet outcomes remain more variable than in children with healthy hearing. This variability cannot be reliably predicted for individual children using age at implant or residual hearing. Development of an artificial intelligence clinical tool to predict which patients will exhibit poorer improvements in language skills may enable an individualized approach to improve language outcomes. Objective To compare the accuracy of traditional machine learning (ML) with deep transfer learning (DTL) algorithms to predict post–cochlear implant spoken language development in children with bilateral sensorineural hearing loss using a binary classification model of high vs low language improvers. Design, Setting, and Participants This multicenter diagnostic study enrolled children from English-, Spanish-, and Cantonese-speaking families across 3 independent clinical centers in the US, Australia, and Hong Kong. A total of 278 children with cochlear implants were enrolled from July 2009 to March 2022 with 1 to 3 years of post–cochlear implant outcomes data. All children underwent pre–cochlear implant 3-dimensional volumetric brain magnetic resonance imaging (MRI). ML and DTL algorithms were trained to predict high vs low language improvers in children with cochlear implants using neuroanatomical features from presurgical brain MRI. Data were analyzed from August 2023 to April 2025. Exposures Cochlear implants. Main Outcomes and Measures The accuracy, sensitivity, and specificity of prediction models based on brain neuroanatomic features using traditional ML and DTL learning. Results Of 278 children, 137 (49.3%) were female, and the mean (SD) age at cochlear implant was 25.7 (18.8) months. DTL prediction models using bilinear attention-based fusion strategy achieved an accuracy of 92.39% (95% CI, 90.70%-94.07%), sensitivity of 91.22% (95% CI, 89.98%-92.47%), specificity of 93.56% (95% CI, 90.91%-96.21%), and area under the curve of 0.98 (95% CI, 0.97-0.99). DTL outperformed traditional ML models in all outcome measures. Conclusions and Relevance The results of this diagnostic study suggest that DTL prediction of language improvement on the individual child level using neuroanatomic features demonstrates greater accuracy, sensitivity, and specificity than traditional ML prediction. DTL was substantially improved by direct capture of discriminative and task-specific information that are advantages of representation learning enabled by this approach vs ML. The results support the feasibility of a single DTL prediction model for language prediction for children served by cochlear implant programs worldwide. Prediction of low improvement may enable targeted early and customized intervention to improve language.