Manual seizure detection in clinical electroencephalography (EEG) is time consuming and requires extensive training. In addition, the seizure origin and spreading pattern is valuable for therapeutic planning but cannot always be manually disambiguated. Prior work in automated seizure detection has focused on engineering new features that better capture the seizure activity. However, these methods ignore crucial information in the data and are not sensitive enough to track the seizure propagation. In this work we introduce a hybrid Probabilistic Graphical Model-Convolutional Neural Network (PGM-CNN) for seizure tracking in multichannel EEG. Our model leverages the power of deep learning for data driven analysis of the raw EEG time series while retaining clinically relevant information through the latent PGM prior. We validate our hybrid model on clinical EEG data from two hospitals with distinct patient populations. Our system achieves better detection performance than baseline methods, which exclusively use PGMs or neural networks.