Electroconvulsive therapy (ECT) is a common clinical tool used for alleviating drug-resistant depression. During ECT, an electrode pair delivers few seconds of repetitive large-amplitude brief electric current pulses onto the patient's scalp that trigger brief ictal seizures in the underlying brain networks, which in turn promote reduction of depression symptoms. ECT therapeutic efficacy depends on the initiation location of a seizure and its propagation patterns. Accordingly, it is very desirable to utilize multichannel EEG recordings to understand how seizures vary with parameters related to electrode configuration (e.g. location, size) and current (e.g. amplitude, polarity). However, a major hurdle to obtaining clear recordings during a seizure (ictal activity, ∼ 100 μv) are the long-lasting (6-8 seconds) very large amplitude (∼ 100 mV) artifacts created by the stimulation electrodes. The disparity in signal-to-noise amplitude during an artifact (∼ 1/1000) constitutes a major challenge to common signal conditioning and artifact removal tools. We here propose a multiple stage algorithm for detecting frequency coherences in hidden ictal EEG signals. The algorithm exploits both temporal information and spatial correlation patterns among affected electrode to remove major artifact components. We present a simulated example to verify the efficacy of our method as well as some preliminary results on actual EEG data.