At a prevalence of almost 1%, potential epileptic seizures manifest a significant health risk for many juvenile patients. Thus, monitoring is essential to set early counteractive measurements to prevent from damage. The sensor-based monitoring systems mainly address epileptic seizures indicated by a change in the muscle tonus but cannot be utilized for patients that show Prévost’s-sign only. To monitor initiating Prévost’s-sign with opened-eyes as critical visual feature, the applicability of deep-learning eye detection systems on night vision images is evaluated in this paper as basis for modelling and classifying the eye state (closed, opened, not visible). A holistic research prototype is presented as proof of concept, showing the applicability of state-of-the-art face detection on night vision images as well as multi-variate feature analysis on Graph segmentation pre-fragmentation, applicable to detect the state of the eye in a robust way. Results show a single frame accuracy in face/eye detection of 73.91% and 94.44% for classification of the opened eyes as indication of a potentially initiating epileptic seizure. The monitoring system is based on a Raspberry computation unit with two ELP night vision cameras attached and a smart phone app for user-interaction and configuration besides on-demand visual monitoring. Future work will show that the single frame detection rate is sufficient for building up a rule-based monitoring state machine at user predefined sensitivity and specificity by analysing the visual content as time-series rather than single images.