A 10.13µJ/Classification 2-Channel Deep Neural Network Based SoC for Negative Emotion Outburst Detection of Autistic...

Area of Science:

• Biomedical Engineering
• Neuroscience
• Integrated Circuit Design

Background:

• Autism Spectrum Disorder (ASD) is characterized by challenges in emotional regulation.
• Non-invasive methods for monitoring emotional states in autistic individuals are crucial for timely intervention.
• Existing neuro-feedback systems often face limitations in power efficiency and hardware footprint.

Purpose of the Study:

• To develop and present an electroencephalogram (EEG)-based, non-invasive, 2-channel neuro-feedback System-on-Chip (SoC).
• To accurately predict and report negative emotion outbursts (NEOB) in autistic patients.
• To achieve area- and power-efficient hardware implementation for on-sensor emotion classification.

Main Methods:

• An area-and-power efficient dual-channel Analog Front-End (AFE) was designed with shared EEG channels to reduce hardware area.
• A customized 4-layer deep neural network (DNN) emotion classification processor was integrated on-sensor.
• The DNN processor utilizes a two-feature vector per channel to minimize area and prevent overfitting.

Main Results:

• The SoC achieved a 30% area reduction due to shared EEG channels in the AFE.
• Excellent AFE performance was demonstrated with an input-referred noise of 0.55 µV_RMS, NEF of 2.71, and crosstalk of -79 dB.
• The 16 mm² SoC, implemented in a 0.18 µm CMOS process, achieved >85% average accuracy in predicting NEOB with a power consumption of 10.13 μJ/classification.

Conclusions:

• The presented SoC offers an efficient and accurate solution for non-invasive detection of negative emotion outbursts in autistic individuals.
• The on-sensor DNN processing minimizes data transmission and enhances real-time responsiveness.
• This technology holds promise for developing advanced neuro-feedback interventions for ASD management.