oa Classification of newborn EEG signals using adaptive time-frequency distributions

Background: EEG signal are widely used for detecting abnormalities such as seizures in newborn babies. EEG signals have non-stationary characteristic, therefore time frequency distributions (TFDs) are a preferred tool for their analysis. The classification performance of most time-frequency (t-f) methods is restricted by the resolution limitation of TFDs. Objective: The main objective of this research is to propose a new time-frequency pattern recognition technique that improves the performance of earlier methods by defining a new high resolution data adaptive time-frequency distribution. Methods: The key steps of the proposed t-f pattern recognition scheme are 1. Transformation of an EEG signal in the t-f domain by using a high resolution TFD. 2. Estimation of the instantaneous frequency (IF) and instantaneous amplitude (IA) of signal components. 3. Extraction of statistical features such as mean, variance, skewness and kurtosis from the estimated IF and IA. 4. Training of a support vector machines using extracted features. The accurate estimation of the IF and IA of signal components is a key step of the proposed t-f pattern recognition technique. Estimation of the IF and IA of signal components depends on the ability of a TFD to resolve closely spaced signal component. In order to overcome the resolution limitation of existing TFDs, we define a new high resolution data adaptive TFD that adapts the direction of its smoothing kernel at each point in the t-f plane based on the direction of energy concentration in the t-f plane. The proposed adaptive TFD out performs other standard methods of t-f analysis in terms of its resolution and instantaneous frequency estimation capabilities. Results: The proposed t-f pattern recognition methodology is applied to detect seizure activity in newborn EEG signals. The classification performance of widely used TFDs such as the extended modified B-distribution, compact support kernel, spectrogram, and proposed adaptive TFD is compared using the leave-one out cross-validation technique. The proposed TFD outperforms other TFDs as well as earlier methods of seizure detection by achieving the total accuracy of 97.5%. Conclusions: A new time-frequency pattern recognition technique for the classification of EEG signals is presented. Results indicate that the performance of the time frequency pattern recognition techniques is sensitive to the resolution performance of TFDs as high resolution TFDs have achieved better classification results.