A low-cost and energy-efficient EEG processor for continuous seizure detection using wavelet transform and AdaBoost

A 16-channel, low-complexity and energy-efficient electroencephalography (EEG) processor for patient-specific seizure detection is presented. The feature extraction algorithm extracts line length features from approximation coefficients from a low-overhead 2-level lifting-based discrete wavelet transform (LWT). Low computational feature vector construction is performed by the concatenation of features extracted from contiguous epochs. A low-complexity binary Adaboost classifier using decision trees is employed in the classification stage. Word length reduction for parameters of the classifier is carried out to make a trade-off between the classification performance and hardware cost. Experimental results demonstrate that our detection algorithm has an average sensitivity, average false alarm rate and latency of 93.8%, 0.16 false alarms/hour and 1 s, respectively. The post-layout simulation in TSMC 65 nm CMOS shows that the proposed processor consumes 0.39 μJ/classification, which is equivalent to 0.006 nJ per input bit.