Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Neural Prediction of Spoken Language Outcomes in Children with Cochlear Implants: A Multi-Center Study
Poster Session A - Sandbox Series, Thursday, October 24, 10:00 - 11:30 am, Great Hall 3 and 4
This poster is part of the Sandbox Series.
Yanlin Wang1, Di Yuan1, Shani Dettman2, Dawn Choo2, Denise Thomas3, Maura E Ryan4, Shimeng Xu5, Nancy M Young3, Patrick C M Wong1; 1Brain and Mind institute, The Chinese University of Hong Kong, 2Department of Audiology & Speech Pathology, The University of Melbourne, 550 Swanston St, Parkville, Victoria 3010 Australia, 3Department of Audiology, Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, Illinois, USA, 4Department of Medical Imaging, Ann & Robert H. Lurie Children's Hospital of Chicago, Chicago, Illinois, USA, 5Department of Molecular Genetics, University of Texas Southwestern Medical Center, Dallas, TX 75390
Cochlear implantation has shown to be an effective treatment method to facilitate spoken language development for children with severe to profound hearing loss. Despite early implantation, spoken language development can be quite variable for children with cochlear implants (CIs). Enrolment into early intervention to enhance spoken language development is effective for children with CIs but can be costly. Being able to predict spoken language development before CIs could facilitate allocation of early intervention, that is a higher dose intervention to children who need it most. Our early studies showed that pre-CI MRI neuroanatomical scans obtained as part of the standard assessment protocol could be used to construct predictive models to forecast language outcomes. However, these studies were restricted to children from one medical center who were learning English. It is important to ascertain whether neural predictive models constructed with data from one medical center who are learning one language can be used to predict the outcomes of children who are from other medical centers who are learning other languages. Our multi-center study addressed this question. A total of 278 children from English, Spanish, and Cantonese language-dominant homes who were diagnosed with congenital/early onset sensorineural hearing loss were recruited from Chicago (English: N=143, Spanish: N=37), Melbourne (English: N=81), and Hong Kong (Cantonese: N=17). All children underwent T1-weighted volumetric neuroanatomical magnetic resonance imaging (MRI) as a part of their pre-CI evaluation. Speech and language abilities were examined before and up to three years post-CIs. The slope of speech and language change was calculated for each child as a measure of improvement, with median split used to categorize children into higher and lower improvement groups. Slice-based deep transfer learning models were constructed where neural features were used to predict improvement (higher vs lower), including models with data from one medical data and models with children learning only one language (English or Spanish). We then tested the generalizability of these models across medical centers and languages. Results showed that deep learning models, particularly the MobileNet model, achieved high predictive performance within specific datasets (AUC: 87.1%, ACC: 89.7%, sensitivity: 94.1%, specificity: 92.2%). However, performance dropped to chance levels when models were tested on different datasets (e.g., Melbourne tested with Chicago data) or across languages within the same site (e.g., Spanish tested with English data in Chicago). When all datasets were combined into a large heterogeneous dataset, the predictive performance remained high (AUC: 0.924, ACC: 87.9%, sensitivity: 88.3%, specificity: 87.6%). This suggests that heterogeneous datasets can facilitate generalization across different sites and languages. In conclusion, using pre-CI neuroanatomical data for predicting language outcomes shows promise for precision care, as evidenced by high accuracy in single homogeneous datasets. Nevertheless, our findings indicate no evidence for cross-site and cross-language generalization within the tested sample sizes, underscoring the necessity for multi-center collaborations to gather larger, more diverse datasets. This approach could enhance the development of models capable of generalizing to new patients from varied backgrounds, optimizing resource allocation for early interventions.
Topic Areas: Computational Approaches, Language Development/Acquisition