A Low-Power, Analog-Integrated, Current-Mode, and Fully Tunable Artificial Neural Network Classifier Architecture for Biomedical Engineering Applications

This work presents a low-power, current-mode analog artificial neural network (ANN) classifier implemented in the TSMC 65-nm CMOS process and evaluated on two biomedical classification tasks. The architecture incorporates analog Sigmoid, Euclidean distance, min and argmin computation blocks to enable fully integrated feature correlation, weight adaptation, and decision-making. All circuits operate in the subthreshold regime with <tex-math notation="LaTeX">± 0.3V </tex-math> supply rails. Post-layout simulations show a worst-case classification accuracy of 91.7% and strong robustness to Monte-Carlo mismatch and Process-Voltage-Temperature variations. Comparison with an equivalent software classifier and prior works validates the modeling and design methodology. The architecture achieves worst-case power consumption below 1311 nW, demonstrating its suitability for energy-constrained edge-AI applications in biomedical monitoring and diagnostics.