Abstract Background and aims Recanalisation in the course of ischemic stroke has been proposed to be associated with a reduction of abnormal slow EEG activity related to disrupted cortical function. EEG assessment has, however, never been performed continuously as a functional monitoring tool during a recanalisation procedure. Here, we aimed to characterize EEG changes during mechanical thrombectomy and to evaluate the ability of conventional EEG metrics and a machine-learning approach to detect recanalization. Methods We studied n=20 subjects with acute anterior circulation stroke due to proximal artery occlusion who underwent continuous EEG monitoring (14 electrodes) during mechanical thrombectomy using classical power ratios (delta- and theta-to-alpha ratios, DAR/TAR) and their respective symmetry indices, as well as a shallow machine learning model with 3 attention layers. Classification performance was assessed within the receiver-operating characteristic (ROC) framework. Results The temporal evolution of power ratios showed considerable variability among subjects, with 7/20 exhibiting an unexpected increase in slow EEG activity immediately after recanalisation. We did not identify any statistically significant clinical predictors of this paradoxical response. While the discriminative performance of power ratios for immediate recanalization detection was poor (AUC 0.55), the machine learning classifier achieved considerably improved performance (AUC 0.7-0.8) at inference times suitable for device-embedded real-time signal processing. Conclusions EEG changes during recanalization are more complex than previously predicted and exhibit considerable inter-subject variability. Reliable immediate detection of recanalisation is not effective with classical power ratios and likely requires using machine-learning techniques. Such AI-based tools may be particularly clinically useful when confirmatory neuroimaging is unavailable after thrombolysis. Conflict of interest Pyrzowski Jan: nothing to disclose; Quirins Marion: nothing to disclose; Zeidan Sinead: nothing to disclose; Mouder Nabila: nothing to disclose; Tahon Anais: nothing to disclose; Sabben Candice: nothing to disclose; Raynouard Igor: nothing to disclose; Piotin Michel: nothing to disclose; Huberfeld Gilles: nothing to disclose; Obadia Michael: nothing to disclose;
Pyrzowski et al. (Fri,) studied this question.