A 1.83 μJ/Classification, 8-Channel, Patient-Specific Epileptic Seizure Classification SoC Using a Non-Linear Support...

Area of Science:

• Biomedical Engineering
• Signal Processing
• Integrated Circuit Design

Background:

• Epilepsy monitoring requires efficient and accurate seizure detection systems.
• Existing solutions often lack integration, portability, or sufficient power efficiency.
• On-chip processing of electroencephalogram (EEG) data presents a challenge for real-time analysis.

Purpose of the Study:

• To develop and present a novel System-on-Chip (SoC) for automated seizure classification in epileptic patients.
• To integrate key components including EEG data acquisition, feature extraction, NLSVM classification, and data storage onto a single chip.
• To optimize the SoC for minimal area and energy consumption while maintaining high performance.

Main Methods:

• Design of a hardware-efficient non-linear support vector machine (NLSVM) classification engine.
• Implementation of a time division multiplexing (TDM)-bandpass filter (BPF) architecture for feature extraction.
• Integration of low-noise, high dynamic range readout circuits with a chopper-stabilized DC servo loop.
• Utilizing a log-linear Gaussian basis function (LL-GBF) NLSVM for linearization and area reduction.

Main Results:

• The developed SoC integrates an NLSVM classifier, feature extraction engine, and 96 KB SRAM for EEG storage.
• The LL-GBF NLSVM classifier achieved a 28.2% area reduction and 0.39 μJ/operation energy consumption.
• Readout circuits exhibited a noise relative to input (RTI) of 0.81 μVrms and an equivalent noise figure (NEF) of 4.0.
• The 5x5 mm² SoC, implemented in a 0.18 μm CMOS process, consumed 1.83 μJ/classification for 8-channel operation.

Conclusions:

• The presented SoC represents a significant advancement in integrated epilepsy monitoring systems.
• The design demonstrates high efficiency in terms of area, energy consumption, and noise performance.
• Verification using the Children's Hospital Boston-MIT EEG database yielded a 95.1% average detection rate with minimal false alarms (0.94%) and 2s latency.
• This integrated solution holds promise for improved wearable devices for epileptic patients.