Low Latency Single-Cycle EOG Classification Using Cascaded ANN & CNN

Electrooculography (EOG) is a noninvasive method to record the motion of the eye from the electrical potential difference between the cornea and retina. This technique, paired with AI, is extensively utilized predominantly in assistive technologies. The complexity of AI-driven EOG analysis scales with the number and type of expected eye movements. This challenge intensifies when the detection of eye movements relies solely on a single EOG cycle. An essential metric for real-time applications is the latency of an AI algorithm, which refers to the delay in making predictions. The latency should be smaller than the average human visual reaction time for any AI-driven system to promptly respond. The majority of EOG classification studies tend to overlook latency. In this experimental study, we overcome this multi-facet problem by proposing cascaded Artificial Neural Network (ANN) and Convolutional Neural Network (CNN) architectures that have latency smaller than average human reaction time. Also, we developed a generalized & robust signal processing method to correctly identify single EOG cycle and deliver high prediction performances with either ANN or CNN. To demonstrate, we successfully classified nine distinct eye movements based only on single cycle EOG signal, with the performance metrics reaching close to 100% for both ANN & CNN, and latency well below the reaction time.