Deep learning models have recently shown great success in classifying epileptic samples from patient EEG recordings. Unfortunately, classification-based methods do not provide a sound mechanism to detect the onset of seizure events. In this work, we propose a two-stage framework that provide an reliable onset detection with a novel task formulation of subsequence clustering. Given an EEG sequence, the framework first learns a set of second-level embeddings with label supervision. It then employs model-based clustering to explicitly models long-term temporal dependencies in EEG sequence and identify a set of subsequences. Epochs within a subsequence share a common cluster assignment (normal or seizure), and state transitions represent successful onset detections. Extensive experiments on three datasets demonstrate the effectiveness of our proposed framework, which can outperform other baselines by 5%-11% in classification improvements and shows accurate detections of seizure onset.