12 May 2020 in IEEE Journal of Solid-State Circuits
23 June 2022 in IEEE Transactions on Biomedical Circuits and Systems
During the past two decades, epileptic seizure detection and prediction algorithms have evolved rapidly. However, despite significant performance improvements, their hardware implementation using conventional technologies, such as Complementary Metal–Oxide–Semiconductor (CMOS), in power and area-constrained settings remains a challenging task; especially when many recording channels are used. In this paper, we propose a novel low-latency parallel Convolutional Neural Network (CNN) architecture that has between 2-2,800x fewer network parameters compared to State-Of-The-Art (SOTA) CNN architectures and achieves 5-fold cross validation accuracy of 99.84% for epileptic seizure detection, and 99.01% and 97.54% for epileptic seizure prediction, when evaluated using the University of Bonn Electroencephalogram (EEG), CHB-MIT and SWEC-ETHZ seizure datasets, respectively. We subsequently implement our network onto analog crossbar arrays comprising Resistive Random-Access Memory (RRAM) devices, and provide a comprehensive benchmark by simulating, laying out, and determining hardware requirements of the CNN component of our system. We parallelize the execution of convolution layer kernels on separate analog crossbars to enable 2 orders of magnitude reduction in latency compared to SOTA hybrid Memristive-CMOS Deep Learning (DL) accelerators. Furthermore, we investigate the effects of non-idealities on our system and investigate Quantization Aware Training (QAT) to mitigate the performance degradation due to low Analog-to-Digital Converter (ADC)/Digital-to-Analog Converter (DAC) resolution. Finally, we propose a stuck weight offsetting methodology to mitigate performance degradation due to stuck $R_{\text{ON}}/R_{\text{OFF}}$ memristor weights, recovering up to 32% accuracy, without requiring retraining. The CNN component of our platform is estimated to consume approximately 2.791 W of power while occupying an area of 31.255 mm 2 in a 22 nm FDSOI CMOS process.
1 October 2021 in IEEE Journal of Solid-State Circuits
An active probe and microstimulator SoC for interfacing with peripheral nerves is presented. It performs 64-channel artifact-tolerant neural recording, cuff imbalance compensation by impedance sensing, and neurostimulation for the closed-loop operation. Each recording channel is a second-order opamp-less $\Delta \Sigma $ ADC that consumes 140 nW and occupies 0.01 mm 2 area in 130 nm CMOS. The single-loop $\Delta \Sigma $ architecture achieves second-order noise shaping with two passive integrators. To the best of our knowledge, this yields the lowest power and area of any second-order $\Delta \Sigma $ ADC and the lowest FOM (fJ/conv. step) of any passive second-order $\Delta \Sigma $ ADC (27 fJ/conv. step). The SoC uniquely performs multi-modal input signal recording: voltage (for neural recording) and current (for impedance sensing) are measured concurrently using frequency multiplexing. The SoC also features a 60 MHz energy-efficient inductive powering link and a 600 MHz RF data communication link. The prototype is validated in vivo in the rat sciatic nerve for electroneurogram (ENG) sensing and the correction of impedance-imbalance in cuff electrodes.
14 December 2020 in IEEE Journal of Solid-State Circuits
The presented 16-channel Δ-modulated neural analog-to-digital converter (ADC) exhibits tolerance to input dc offsets of any value, up to the supply voltage. It employs a dynamic differential-difference comparator architecture with a super-GΩ input impedance ensuring negligible gate-leakage current and well-matched differential inputs resulting in more than 78 dB of rejection of common-mode signals and artifacts. The all-digital nature of the presented Δ -ADC enables sampling of input signals at high oversampling ratios (OSRs) making the front-end immune to stimulation artifacts with differential amplitudes up to a limit that is scalable by the OSR (e.g., 10 mVPP at OSR = 10 k). Moreover, it allows the Δ-ADC to linearly scale down the power consumption required by the application's recording bandwidth. The oversampled Δ-ADC achieves an effective number of bits (ENOB) of 9.7-bit and 2.6-μVRMS integrated input-referred noise over 1 Hz to 500-Hz bandwidth. It uses no large passives or statically biased circuits, such as operational amplifiers (Opamps) saving both channel area (0.011 mm²) and power consumption (0.99 μW). Experimentally measured results validate the key features of the design and include in vivo recordings in freely moving guinea pigs. The fabricated prototype system-on-a-chip (SoC) hosts an array of 16-channel neural-ADC with in-channel digitally programmable high-compliance current-mode biphasic stimulators as well as wireless circuitry for data communication and power/command reception.
24 March 2020 in IEEE Journal of Solid-State Circuits
A fully integrated 130-nm CMOS 12-channel biofouling-resistant potassium-selective brain neurochemistry impedance spectrum analyzer is presented. Each 0.004 mm 2 channel is an amperometric readout circuit composed of an opamp-less delta-sigma-modulated analog-to-digital converter (ADC) that consumes only 50 nW of power from a 0.6-V supply. The latter is analytically proven to be approximately equivalent to the former under conditions that are typical for the biochemical microsensors. The circuit also includes a low-power multiplierless frequency response analysis (FRA) unit that performs bioimpedance extraction. The channel achieves 1 pA current sensitivity and attains 50.3-dB signal-to-noise-and-distortion ratio (SNDR). The impedance analyzer achieves a dynamic range of 1 pA to 20 nA and a 5-kHz frequency scan range. The 2 mm × 1 mm die has been validated in vivo in real-time multisite potassium sensing in the rodent brain using gold microelectordes implemented on a polyimide substrate.
13 November 2023 in IEEE Transactions on Biomedical Circuits and Systems
This paper presents an innovative, minimally invasive, battery-free, wireless, peripheral nervous system (PNS) neural interface, which seamlessly integrates a millimeter-scale, fascicle-selective integrated circuit (IC) with extraneural recording and stimulating channels. The system also incorporates a wearable interrogator equipped with integrated machine-learning capabilities. This PNS interface is specifically tailored for adaptive neuromodulation therapy, targeting individuals with paralysis, amputation, or chronic medical conditions. By employing a neural pathway classifier and temporal interference stimulation, the proposed interface achieves precise deep fascicle selectivity for recording and stimulation without the need for nerve penetration or compression. Ultrasonic energy harvesters facilitate wireless power harvesting and data reception, enhancing the usability of the system. Key circuit performance metrics encompass a 2.2 μVrms input-referred noise, 14-bit ENOB, and a 173 dB Schreier figure of merit (FOM) for the neural analog-to-digital converter (ADC). Additionally, the ultra-low-power radio-frequency (RF) transmitter boasts a remarkable 1.38 pJ/bit energy efficiency. In vivo experiments conducted on rat sciatic nerves provide compelling evidence of the interface's ability to selectively stimulate and record neural fascicles. The proposed PNS neural interface offers alternative treatment options for diagnosing and treating neurological disorders, as well as restoring or repairing neural functions, improving the quality of life for patients with neurological and sensory deficits.