Modular DR- and CMR-Boosted Artifact-Resilient EEG Headset With Distributed Pulse-Based Feature Extraction and Neuro-Inspired Boosted-SVM Classifier

Abstract

This article presents the design, development, and experimental testing of a flexible, modular electroencephalography (EEG) headset for long-term epilepsy monitoring and individualized treatment. System- and circuit-level techniques are employed to improve energy efficiency while ensuring high-quality EEG recordings and accurate seizure detection. The wearable prototype includes digital active electrodes (DAEs) for high-dynamic-range (DR) recording, motion artifact removal (MAR), and feature extraction (FE), along with a central backend (BE) for patient-specific classification, wireless connectivity, and common-mode rejection (CMR) boosting. DAEs communicate through a time-shared data bus, minimizing wires and enabling flexible electrode placement, maximizing system scalability. Each DAE enhances recording quality with: 1) calibrated CMR boosting (>80-dB common-mode rejection ratio (CMRR) with 1- $M\Omega $ AE-to-AE mismatch); 2) SC notch filtering for power-line noise; 3) real-time electrode-tissue impedance (ETI) measurement for MAR and dc correction; and 4) an autoranging mechanism with 17-dB DR enhancement. In-AE FE cuts AE-to-BE communication power by 99.1%, while pulse-based frequency sampling reduces FE power by 92.1%. A neuromorphic multiplier-less adaptively boosted support vector machine (SVM) maintains high detection accuracy with 97.6% less classification power than conventional designs. The chip was implemented in 180-nm CMOS, and the wearable system components (DAE, wireless, and CMR boards) were miniaturized. Experimental testing showed IIRN ( $0.64~\mu V_{\text {rms}}$ , 0.5–100 Hz), adjustable gain/bandwidth, DR (80 dB), SNDR (74.5 dB), and CMRR (89.2 dB without mismatch, >80 dB with mismatch). Measurement results also confirm the system’s effectiveness in motion artifact estimation and removal. In vivo measurements demonstrate the system’s efficacy and latency in detecting neurologically relevant events. Seizure detection results (96.4% sensitivity, 0.41 FPR, 1-s latency, zero SRAM usage) on prerecorded EEG data from 21 patients are also reported. The system is compared to state-of-the-art EEG recording and seizure detection systems, highlighting its advantages.