A data-driven modeling approach to stochastic computation for low-energy biomedical devices.

Area of Science:

• Biomedical Engineering
• Signal Processing
• Hardware Reliability

Background:

• Low-power devices for physiological signal analysis enable closed-loop systems.
• Hardware errors in ultra-low-power platforms are a significant limitation.

Purpose of the Study:

• To demonstrate data-driven methods for modeling and overcoming hardware errors in physiological signal processing.
• To achieve robust performance in low-power biomedical devices despite hardware faults.

Main Methods:

• Data-driven modeling of physiological signals (EEG, ECG) and hardware error sources (SRAM bit-cell errors, logic-gate stuck-at faults).
• Synthesis of EEG-based seizure detection and ECG-based arrhythmia-beat classification to logic-gate implementation.
• Evaluation using patient data from CHB-MIT and MIT-BIH databases.

Main Results:

• Data-driven methods effectively model and mitigate prominent hardware error sources with minimal overhead.
• Achieved performance comparable to error-free hardware even at high fault rates (SRAM up to 0.5, logic up to 7x10^-2).
• Maintained high accuracy despite computational bit error rates reaching 50%.

Conclusions:

• Data-driven approaches offer a viable solution to hardware errors in ultra-low-power biomedical devices.
• This methodology enhances the reliability of real-time physiological monitoring and closed-loop therapeutic systems.