Machine Learning-Enabled Prediction of Speech Perception Improvement Based on Diffusion Tensor Imaging of Young Cochlear Implant Candidates

Abstract

AbstractPrediction of improvement in speech perception after cochlear implantation (CI) is clinically important to optimize pediatric habilitation. Conventional methods using non-neural measures do not permit accurate prediction on the individual child level. In this study, we investigate whether white matter patterns detected by diffusion tensor imaging (DTI) magnetic resonance imaging (MRI) predict speech perception improvement. Pre-surgical DTI of CI candidates was compared to matched normal-hearing (NH) children to determine cortical regions affected by hearing impairment. Speech Recognition Index in Quiet was measured at baseline and 6 months post implantation to compute improvement in speech perception. Machine learning prediction of speech perception improvement (high or low) was performed using non-imaging and DTI white matter characteristics of whole, affected and unaffected brain. Affected and unaffected white matter regions were determined by comparing DTI multi-voxel pattern similarity maps of white matter integrity indices between CI and NH. Abnormal white matter patterns were found throughout brain of CI candidates. Prediction of 6-month post-CI improvement accuracy, sensitivity and specificity using unaffected regions (0.86, 0.91, 0.80, respectively) and whole brain white matter (0.85, 0.91, 0.80, respectively) yielded similar results, and were more predictive than regions affected by hearing impairment (0.72, 0.74, 0.70, respectively) or non-imaging features (0.67, 0.55, 0.78, respectively). Findings support that presurgical neural white matter pathways, especially in the association auditory and cognitive regions not affected by auditory deprivation, play a critical role in speech development after CI and are more predictive of outcome than traditional non-neural variables such as age at implant.