Transcranial magnetic stimulation language mapping analysis revisited: Machine learning classification of 90 patients reveals distinct reorganization pattern in aphasic patients

Repetitive TMS (rTMS) allows to non-invasively and transiently disrupt local neuronal functioning. Its potential for mapping of language function is currently explored. Given the inter- individual heterogeneity of tumor impact on the language network and resulting rTMS derived functional mapping, we propose to use machine learning strategies to classify potential patterns of functional reorganization. We retrospectively included 90 patients with left perisylvian glioma tumors, world health organization (WHO) grade II-IV, affecting the language network. All patients underwent navigated rTMS language mappings. The severity of aphasia was assessed preoperatively using the Berlin Aphasia Score (BAS), which is adapted to the Aachener Aphasia Test (AAT). After spatial normalization to MNI 152 of all rTMS spots, we calculated the error rate (ER) in each cortical area by automated anatomical labeling parcellation (AAL) and used support vector machine (SVM) as a classifier for significant areas in relation to aphasia. 29 of 90 (32.2%) patients suffered from aphasia. Univariate analysis revealed 11 perisylvian AVOIs' ERs (eight left, three right hemispheric) that were significantly higher in the aphasic than non-aphasic group (p < 0.05), depicting a broad, bihemispheric language network. After feeding the significant AVOIs into the SVM model, it showed that additional to age (w = 2.95), the ERs of right Frontal_Inf_Tri (w = 2.06) and left SupraMarginal (w = 2.05) and Parietal_Inf (w= 1.80) contributed more than other features to the model. The model's sensitivity was 89.7%, the specificity was 82.0%, the overall accuracy was 81.1% and AUC was 88.7%. Our results demonstrate an increased vulnerability of the right inferior frontal gyrus to rTMS in patients suffering from aphasia due to left perisylvian gliomas. This confirms a functional relevant involvement of the right frontal area in relation to aphasia. While age as a feature improved our SVM model the most, the tumor location feature didn't affect the SVM model. This finding indicates that general tumor induced network disconnection is relevant to aphasia and not necessarily related to specific lesion locations. Additionally, our results emphasize the decreasing potential for neuroplasticity with age.