DirichNet Model for Detection of TMS-Induced Speech Errors in Patients Undergoing Epilepsy Surgery

Abstract

Presurgical mapping of critical cortical regions is crucial for effective surgical planning and minimizing post-operative deficits. Transcranial magnetic stimulation (TMS) provides a non-invasive method for identifying critical brain areas involved in language and speech in neurosurgery patients. However, integrating TMS maps into surgical planning is complicated by inconsistent and subjective error classification criteria across multiple centers, leading to false positives and false negatives. To address this challenge, we introduce DirichNet, an automated deep learning model well suited for objective and reliable detection of TMS-induced speech errors. DirichNet advances the base-line SincNet by incorporating periodic Dirichlet kernels, which improves synchronization with TMS speech data. Experiments with both odd and even kernels show that DirichNet significantly outperforms the other models in accurately identifying the errors. This pilot study is the first to provide automated detection of TMS-induced speech errors, reducing review time for raters and ensuring objective classification.