Teager–Kaiser Energy Operator for the Analysis of Brain Signal Dynamics

Abstract

Brain signals such as Electroencephalography (EEG) can be effectively modeled as Amplitude–Frequency Modulated (AM–FM) signals, reflecting the intrinsic oscillatory and dynamic nature of neural activity. In this representation, the instantaneous amplitude and frequency capture physiologically meaningful information about energy variations and neuronal synchrony. However, the extraction of such components from EEG is challenging due to the signal’s non-linear, non-stationary, and noise-prone characteristics. In this Thesis, we investigate the Teager–Kaiser Energy Operator (TKEO) as a non-linear tool for demodulating EEG signals and analyze their energy dynamics across cognitive and clinical EEG paradigms.

To accommodate the narrowband nature of the operator, a Gabor filterbank is first applied to isolate canonical frequency bands. The Discrete Energy Separation Algorithm (DESA-1) is then employed to extract the amplitude envelope and the instantaneous frequency components within each sub-band, enabling a detailed analysis of amplitude–frequency modulated dynamics. From both the Teager-Kaiser energy and the demodulation components, we derive a comprehensive set of energy-based descriptors. These features are systematically evaluated across three representative EEG classification tasks—Emotion Recognition, Motor Imagery, and Epilepsy Detection—and compared with conventional energy and spectral-domain features, under both Subject-Dependent and Subject-Independent settings to assess generalization performance.

Experimental results demonstrate that TKEO-based features yield higher classification performance in Motor Imagery and Epilepsy Detection, while achieving comparable results in Emotion Recognition. Furthermore, feature fusion across frequency bands enhances overall performance, emphasizing the narrowband sensitivity of TKEO-based descriptors. Analysis of the filterbank configuration shows that increasing the number of filters improves performance in most cases, as finer spectral decomposition enables TKEO to capture subtle energy fluctuations that would otherwise be averaged out. Beyond quantitative improvements, the extracted features retain physiological interpretability, directly reflecting the instantaneous energy and frequency modulations underlying neural activity.