To assess the effectiveness of machine learning in automating the prediction of vestibular abnormalities after cochlear implantation (CI) in patients with sensorineural hearing loss (SNHL), with the goal of developing a practical model that can accurately predict long-term vestibular function outcomes and identify associated risk factors.

Clinical data, including imaging, vestibular evoked myogenic potentials (VEMPs), and auditory information, were collected from patients with sensorineural hearing loss (SNHL) before and after CI. The decision tree algorithm was employed to address missing values and screen pre-CI clinical features. Six machine learning methods were subsequently utilized to predict the relationships between the extracted features and post-CI vestibular dysfunction. The best-performing method determined the ranking of feature importance, which was regarded as risk factors for predicting symptoms and VEMPs results after CI.

Logistic regression models effectively predicted both post-CI vestibular dysfunction and abnormal cervical VEMP (cVEMP), with accuracies of 80% and 78%, respectively. The relative importance of the features, in descending order, was as follows: cVEMP latency, cVEMP amplitude, and residual hearing threshold. Moreover, the support vector machine (SVM) model attained an accuracy of 88% in predicting abnormal ocular VEMP (oVEMP) post-CI. For the SVM model, the feature importance ranking was as follows: oVEMP latency, oVEMP amplitude, and residual hearing threshold.

This study successfully leverages machine learning techniques, specifically support vector machines (SVM) and logistic regression models, to predict the impact of CI on vestibular function. These predictive models provide valuable insights for presurgical planning and decision-making in CI procedures. Moreover, the findings highlight the critical risk factors associated with vestibular dysfunction, offering a robust reference for guiding vestibular rehabilitation strategies.