Bio-Inspired Adaptive Signal Propagation Networks For Enhanced Pattern Recognition In Biomedical Signal Processing Applications

Abstract

However, the biomedical signals like ECG and EEG are non-stationary, noisy, andpatient-specific; yet, the majority of pattern recognition pipelines for deep learning utilize pre-defined, non-biological signal propagation pathways that do not change after being learned in one phase and used the same way regardless of the presence of noise or the person who produces the signal. In this study, we propose a Bio-Inspired Adaptive Signal Propagation Network (BI-ASPN) that employs adaptive threshold leaky integrate and fire spike encoder, recurrent spiking reservoir mimicking the liquid state machine architecture, and spike-timing-dependent plasticity (STDP) to constantly adapt the selected propagation pathways through the reservoir depending on the currently processed noise properties. The lightweight readout classifier works based on spike patterns in the reservoir system, and a homeostatic re-adaptation process is initiated whenever there is a considerable change in input statistics in comparison to what the current architecture of the pathway has been optimized for. The algorithm is validated using the ECG beat classification dataset derived statistically from the PhysioNet MIT-BIH Arrhythmia Database, and four different degraded scenarios: clean signal, noisy signal, highly noisy signal, and cross-subject generalization. In the context of the cross-subject scenario, the suggested BI-ASPN model maintains 88.2 percent accuracy, whereas the other two baselines achieve 71.6 percent and 76.4 percent accuracy respectively. It shows that biologically plausible adaptive signal propagation mechanism can lead to a robust biomedical pattern recognition system.