Emotional states involving anger, hostility, anxiety and stress have been associated with an increased risk of cardiovascular disease. Online emotion recognition has achieved little attention in the literature in comparison to offline approaches. We present both online and offline methods to identify anger based on EMG data. In the offline method, the Hilbert-Huang transform is used to extract energy features from different time-frequency blocks. This approach achieves an overall classification accuracy of 87.5%. We also develop a novel online method combining machine learning with the tracking of a single parameter for anger detection. Here, band energy is calculated within time windows, and is continuously adjusted based on classified peak amplitudes. Although this technique has a lower classification accuracy than the offline method, it is quite promising as it is well-suited for wearable monitoring and long-term stress management.