Generalization of data augmentation to reduce the number of epochs to average

Abstract

The waveform derived from averaging multiple EEG signal recordings during stimulation represents an Event-Related Potential (ERP). When sensory stimuli are employed, the resulting potentials are termed Evoked Potentials (EPs). EPs find applications across diverse domains of research and clinical settings, serving as a valuable tool in neuroscience and medicine due to their versatility in offering objective insights into brain function. However, the conventional signal averaging method used to extract EPs has inherent limitations, such as the necessity for numerous trials to ensure reliability and maximize Signal-to-Noise Ratio (SNR). This demands additional time for data recording and processing. Moreover, the reliability of recorded responses may be compromised due to the subject's habituation to the stimulus. To address these challenges, this study aims to enhance SNR in EP extraction by employing data augmentation, thereby reducing the number of records needed for averaging. The proposed method demonstrates a notable improvement of approximately 9.77 ± 2.65 dB compared to traditional signal averaging with the same number of records. This study concludes that judicious data augmentation enables enhanced SNR estimates without the requirement for extensive new recordings.