In theory ML or simpler statistical models like ICA should be able to separate those signals from each other (if they are separable) given the training data contains measurements from all scenarios (moving eyes, muscles, sweating, reacting to different stimuli etc.).