An Intelligent Multimodal Healthcare IC with On-Chip Computing Capabilities for Miniaturized Wearable Devices

Abstract

Recent advanced wearable healthcare devices provide multimodal bio signal sensing functionality including electrocardiogram (ECG), electroencephalogram (EEG), electrooculogram (EOG) photoplethysmogram (PPG), bioimpedance (BioZ), and galvanic skin response (GSR), to capture important physiological information for comprehensive health monitoring such as cardiovascular disease (CVD) and mental health (MH) applications. These multimodal devices can incorporate electrical stimulation function such as transcutaneous vagus nerve stimulation (tVNS) for biofeedback purposes as well. Alongside multimodal sensing, in-sensor artificial intelligence (AI) capabilities have become essential for real-time signal analysis and diagnosis. The integration of multimodal front-end circuits and AI processing at the integrated circuit (IC) level enables device miniaturization, low-power operation, and continuous monitoring, providing effective ambulatory health assessment. This IC based solution addresses critical challenges of conventional remote server-based systems including excessive power consumption from data transmission, long latency in time-critical medical events, and privacy concerns associated with transmitting personal health data over wireless networks. This work presents the intelligent multimodal biomedical readout IC (ROIC) design which integrates various multimodal biomedical channels and an analog computing-based heart rate (HR) extraction and arrhythmia classifier unit. The multimodal analog front ends (AFE) were designed to address each signal modality’s sensing requirements. The biopotential (ExG) channel was designed with low noise of 0.22μVrms and high DC input impedance of 2.5GΩ for small EEG signal detection which is based on the chopper stabilized capacitively coupled instrumentation amplifier (CS-CCIA) architecture. In addition, ±420mV electrode DC offset (EDO) cancellation range was also achieved to address electrode tissue impedance (ETI) variation. For GSR sensing, wide dynamic range (DR) of 130dB was achieved and a resource efficient ECG hybrid GSR recording scheme is presented. For multimodal feasibility, PPG and BioZ channels with 125dB and 145dB DR, and other essential blocks such as Successive approximation register (SAR) analog-to-digital converter (ADC), low-dropout (LDO) regulator, and serial peripheral interface (SPI) blocks were integrated. Finally, a 10V supply tVNS channel with programmable current and stimulation waveform was also integrated. Additionally, the IC supports on-chip HR extraction and arrhythmia classification functionality. The on-chip HR extraction is based on the adaptive analog peak detector circuit and an R-R interval (RRI) counter with normalization option. The on-chip classifier consists of a multi-domain convolutional neural network (MD-CNN) first stage and subsequent analog ternary neural network (TNN) stages. The analog multiply-and accumulate (MAC) block in MD-CNN is based on a time-domain current integrating digital to analog converter (CI-DAC), which consists of a single PMOS current source and weight capacitors per channel, designed to minimize the device mismatch and achieve wide swing range. Additionally, this CI-DAC based MAC operation can incorporate both analog and digital inputs and can be reconfigured according to the target application. The on-chip computation unit achieved sensitivity and specificity of 92.80%/96.36% for three-label arrhythmia classification using MIT-BIH database. The proposed IC was fabricated in a 0.18μm BCD process and integrated into miniaturized Bluetooth low energy (BLE) based wireless biomedical devices in two different types of wearable form factors including behind-the-ear (BTE) device and chest patch device. Both devices are based on the adhesive patch-type configuration, which enables reliable signal monitoring based on the body conformability, adhesiveness, and electrode quality. The chest patch device is targeted for CVD monitoring, where the ECG, PPG, phonocardiogram (PCG), respiration, motion and pulse arrival time (PAT) based BP measurements are supported. The energy efficient on-chip HR extraction was demonstrated compared to conventional PC based digital signal processing (DSP). Additionally, the patch heating functionality and temperature monitoring function for thermal-switching patch adhesion control are also featured. The BTE device supports EEG, ECG, PPG, GSR, BP, motion recording, and auricular tVNS (taVNS) functionality for MH monitoring applications. It supports convenient hands-free and unobtrusive recording for MH related BP or EEG recording in daily life monitoring. The system-level feasibility was experimentally verified through various in-vivo bio signal monitoring. Finally, the BTE device measured BP for comparison with a conventional cuff-based device and measured in vivo stress using EEG under virtual reality (VR) conditions, validating the proposed IC and system for both CVD and MH monitoring applications. Keywords: Multimodal biomedical system, readout integrated circuit (ROIC), electrode artifact, chopper stabilized capacitive coupled instrumentation amplifier (CS-CCIA), low noise, input impedance boosting, EEG, ECG, PPG, GSR, taVNS, stress measurement, pulse arrival time (PAT), heart rate (HR), blood pressure (BP), analog computing, multi-domain convolution neural network (MD-CNN), ternary neural network (TNN), behind-the-ear (BTE), patch-type wearable device.