Neural Prediction of Spoken Language Improvements in Children with Cochlear Implants

Objective This study aims to construct neural predictive models to forecast post-CI spoken language improvements in children with hearing loss and to evaluate whether these models are language- and center-specific.Methods A total of 278 children with hearing loss underwent magnetic resonance image (MRI) examinations and completed speech and language assessments both before and after the implants. We utilized deep transfer learning algorithms with pre-CI neuroanatomical features to predict post-CI spoken language development in children enrolled from 2009 to 2022, with 3-year follow-up. Results We found that pre-CI MRI brain data can forecast spoken language development up to 36 months post-CI. Evidence of within-center and within-language prediction was consistent across different centers. MobileNet model exhibited the best performance with an accuracy (ACC) of 89.74% (95% CI, 89.39%-90.10%), sensitivity of 87.09% (95% CI, 86.17%-88.00%), specificity of 92.20% (95% CI, 90.98%-93.42%), and the area under the receiver operating characteristic curve (AUC) of 0.896 (95% CI, 0.893-0.900). However, cross-dataset generalization, even within the same center, could not be achieved with our current sample (e.g., ACC: 50.27% (95% CI, 47.62%-53.76%), sensitivity: 36.89% (95% CI, 0%-93.88%), specificity: 63.95% (95% CI, 6.43%-100%), and AUC: 0.499 (95% CI, 0.467-0.532)). When all the datasets were combined, the predictive performance remained high (ACC: 87.94% (95% CI, 87.28%-88.59%), sensitivity: 88.33% (95% CI, 97.18%-89.48%), specificity: 87.56% (95% CI, 86.12%-89.00%), and AUC: 0.879 (95% CI, 0.873-0.886)).Conclusions The generalization of the neural predictive model across different centers and languages appears to be feasible and effective with a larger and more representative dataset.