A semi‐supervised Support Vector Machine model for predicting the language outcomes following cochlear implantation based on pre‐implant brain fMRI imaging

INTRODUCTION: We developed a machine learning model to predict whether or not a cochlear implant (CI) candidate will develop effective language skills within 2 years after the CI surgery by using the pre‐implant brain fMRI data from the candidate. METHODS: The language performance was measured 2 years after the CI surgery by the Clinical Evaluation of Language Fundamentals‐Preschool, Second Edition (CELF‐P2). Based on the CELF‐P2 scores, the CI recipients were designated as either effective or ineffective CI users. For feature extraction from the fMRI data, we constructed contrast maps using the general linear model, and then utilized the Bag‐of‐Words (BoW) approach that we previously published to convert the contrast maps into feature vectors. We trained both supervised models and semi‐supervised models to classify CI users as effective or ineffective. RESULTS: Compared with the conventional feature extraction approach, which used each single voxel as a feature, our BoW approach gave rise to much better performance for the classification of effective versus ineffective CI users. The semi‐supervised model with the feature set extracted by the BoW approach from the contrast of speech versus silence achieved a leave‐one‐out cross‐validation AUC as high as 0.97. Recursive feature elimination unexpectedly revealed that two features were sufficient to provide highly accurate classification of effective versus ineffective CI users based on our current dataset. CONCLUSION: We have validated the hypothesis that pre‐implant cortical activation patterns revealed by fMRI during infancy correlate with language performance 2 years after cochlear implantation. The two brain regions highlighted by our classifier are potential biomarkers for the prediction of CI outcomes. Our study also demonstrated the superiority of the semi‐supervised model over the supervised model. It is always worthwhile to try a semi‐supervised model when unlabeled data are available.