In-sensor neural network for high energy efficiency analog-to-information conversion

Area of Science:

• Biomedical Engineering
• Signal Processing
• Computer Science

Background:

• Electrocardiograph (ECG) signals are crucial for diagnosing various health conditions.
• Traditional methods require significant data transmission, leading to high power consumption and limited real-time analysis.
• In-sensor processing of biomedical data is essential for efficient remote monitoring.

Purpose of the Study:

• To develop an on-chip analog-to-information conversion technique for local ECG signal processing.
• To reduce radio frequency transmission of ECG data by over three orders-of-magnitude.
• To demonstrate the application of this technique for early sepsis detection with high accuracy and energy efficiency.

Main Methods:

• Utilized analog hyper-dimensional computing based on the reservoir-computing paradigm for in-sensor ECG analysis.
• Implemented a nonlinear reservoir kernel followed by an artificial neural network for signal processing.
• Prototyped test-chips using 65 nm CMOS technology for validation.

Main Results:

• Achieved reduction in radio frequency transmission by more than three orders-of-magnitude.
• Demonstrated state-of-the-art accuracy in sepsis onset detection.
• Significantly improved energy efficiency, reducing sensor power by [Formula: see text].

Conclusions:

• The proposed on-chip analog-to-information conversion technique enables efficient, local processing of ECG signals.
• This technology offers a promising solution for low-power, real-time health monitoring and early disease detection, such as sepsis.
• The successful prototyping validates the feasibility and performance of the developed system.