An application of Integrated Gradients for the analysis of the P3b event-related potential on a convolutional neural network trained with superlets

Event-related potentials (ERPs) are specific transient fluctuations in the brain’s electrical field induced by the presentation of visual or auditory stimuli. They are time-locked and can help analyse neural activity in many applications. One of these, the P3b, elicited in the decision-making process, unlike other ERPs, is not linked to the physical characteristics of a stimulus but rather to a person’s reaction to it. P3b ERPs can be evoked when completing a task with the oddball paradigm and have only slightly different waveforms when the target and non-target stimuli are presented. Identifying differences in these waveforms can help investigate dysfunctions in sensory and cognitive processing, among other applications. However, such identification is not straightforward because these differences are subtle and neural data contains many artefacts. Recently, deep learning has been used in addition to traditional methods to automatically learn the high-level features associated with target and non-target stimuli, discarding artefacts and supporting their discrimination. Unfortunately, even if powerful in creating discriminative models, they are regarded as black boxes because their inferential capacity is opaque and obscured. This research builds on this gap and explores an application of Integrated Gradients, a method developed within eXplainable Artificial Intelligence, to interpret a convolutional deep neural network, trained with superlets, specific time-frequency super-resolution of single-channel EEG signals, for discriminating target and non-target neural responses for an oddball task.