Acta Electronica Malaysia (AEM)

The continuous monitoring of electrocardiogram (ECG) signals is essential for the early detection of cardiacarrhythmias. However, the power and area constraints of implantable and wearable devices demand ultra￾low-power and highly efficient processing solutions. Conventional architectures, which digitize raw data athigh rates for digital signal processing (DSP), are power-prohibitive for always-on operation. This paperpresents a mixed-signal System-on-Chip (SoC) that leverages analog-compute-in-memory (ACIM) to performfeature extraction and classification directly within the analog domain, minimizing energy-intensive datamovement and high-resolution analog-to-digital conversion (ADC). The proposed SoC integrates a front-endanalog Haar wavelet transform (HWT) engine, an ACIM-based binary classifier, and a 6-bit successiveapproximation register (SAR) ADC on a single die. The HWT is implemented using a switched-capacitornetwork that computes wavelet coefficients at the circuit level, extracting key time-frequency features fromthe ECG waveform. The classifier is built around a non-volatile analog memory array (using resistive RAM(ReRAM) cells) that stores learned weights for a support vector machine (SVM) model, enabling matrix￾vector multiplication through Ohm’s and Kirchhoff’s laws in a single step. Fabricated in a 65 nm CMOSprocess, the entire system operates at a supply voltage of 0.5 V. Experimental results on the MIT-BIH Arrhythmia database show that the analog front-end achieves 28.6 dB of noise suppression. The ACIM classifier consumes only 28 nJ per classification and achieves an accuracy of 97.8% in detecting ventricular ectopic beats (VEBs), outperforming a comparable digital implementation running on a Cortex-M0 microcontroller. The total power consumption is 28 nW at a heart rate of 72 bpm, making it suitable for long￾term, self-powered wearable and implantable ECG monitors.