MulPi: A Multi-class and Patient-independent Epileptic Seizure Classifier with Co-designed Input-stationary Computing-in-SRAM

Area of Science:

• Biomedical Engineering
• Neurology
• Computer Science

Background:

• Epilepsy affects over 70 million people globally, with unprovoked seizures posing significant risks.
• Automated seizure detection and prediction are crucial for improving patient quality of life and preventing fatal outcomes.
• Existing seizure classification chips often lack the multi-class and patient-independent capabilities required for authentic detection due to resource constraints.

Purpose of the Study:

• To introduce the first efficient multi-class, patient-independent (MulPi) chip for automated seizure classification.
• To develop a co-designed system integrating a lightweight convolutional neural network (CNN) with efficient hardware.
• To address the limitations of current implementable chips in terms of accuracy and resource efficiency for seizure detection.

Main Methods:

• Developed a 5-layer convolutional neural network (CNN), named MulPiCNN, employing advanced training techniques for optimized accuracy and reduced complexity.
• Designed an SRAM-based chip utilizing computing-in-memory (CIM) architecture for enhanced efficiency.
• Implemented a novel Input-Stationary, Row-wise CIM (ISRW-CIM) with a 2T-Hadamard product unit (HPU) to overcome SRAM CIM challenges and save resources.

Main Results:

• The fabricated MulPi chip achieved high performance metrics: 98.5% sensitivity and 99.2% specificity.
• The chip demonstrates superior efficiency, with classification completed in 0.12 seconds and occupying a small area of 0.348 mm².
• The ISRW-CIM architecture proved effective in resource-saving and addressing SRAM CIM limitations.

Conclusions:

• The MulPi chip represents a significant advancement in automated seizure detection technology.
• This efficient, co-designed solution simultaneously fulfills multi-class and patient-independent classification requirements.
• The developed chip offers a promising pathway towards enabling seizure-free lives for epilepsy patients through accurate and efficient detection.