Neuro-ElectroFusion based Intelligent Embedded System for Adaptive Biomedical Signal Enhancement

Signal enhancement of biomedical signals is an important area of real-time biomedical monitoring, particularly in resource-constrained embedded systems. Conventional filtering methods tend to confuse noise and critical signal changes whereas pure neural methods require a lot of computation which is inapplicable to low-power wearable devices. This paper presents Neuro-ElectroFusion which is an intelligent embedded system that provides a lightweight neural processing with rule-based signal correction that would provide adaptive improvement of multi-modal biomedical signals, including EEG, ECG, and EMG. The system is based on a quantized 1D-CNN-BiLSTM architecture and a fuzzy logic unit that dynamically denoises and reconstructs signals, depending on the context and physiological state. An entropy-based event-driven sampling scheme consumes considerably less energy and latency and has no fidelity trade-offs. PhysioNet and MIT-BIH performance On PhysioNet and MIT-BIH data sets, the average SNR improvement of 13.7 dB, the reduction in RMSE of 73.4% and the real-time latency of the inference of 16.3 ms per signal window were obtained. The proposed system recorded a high accuracy of 97.68% in various noise levels, which was higher than the traditional DSP-based and TinyML models. Moreover, context-conscious Q-learning was capable of providing adaptive signal prioritization and it was noteworthy that over the baseline systems, context-aware Q-learning improved energy efficiency by 17.6%.