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

Area of Science:

• Neuroscience
• Artificial Intelligence
• Speech Language Pathology

Background:

• Cochlear implants significantly improve spoken language in children with severe to profound sensorineural hearing loss.
• Language outcomes after cochlear implantation are highly variable and difficult to predict.
• Current prediction methods do not reliably identify children who will experience poorer language development.

Purpose of the Study:

• To compare the predictive accuracy of traditional machine learning (ML) and deep transfer learning (DTL) algorithms.
• To predict post-cochlear implant spoken language development in children using neuroanatomical features from brain MRI.
• To classify children as high or low language improvers.

Main Methods:

• A multicenter diagnostic study enrolled 278 children with cochlear implants across the US, Australia, and Hong Kong.
• Presurgical 3D volumetric brain MRI scans were used to extract neuroanatomical features.
• ML and DTL algorithms were trained to predict high vs. low language improvement.

Main Results:

• DTL models achieved 92.39% accuracy, 91.22% sensitivity, and 93.56% specificity.
• DTL significantly outperformed traditional ML models across all outcome measures.
• The area under the curve for DTL was 0.98, indicating high predictive performance.

Conclusions:

• DTL demonstrates superior accuracy, sensitivity, and specificity for predicting individual language improvement post-cochlear implant compared to traditional ML.
• DTL's ability to capture discriminative information via representation learning offers advantages.
• A single DTL model could be feasible for global application, enabling personalized interventions for improved language outcomes.