Aims: A biologically plausible, shallow spiking neural network (SNN) is proposed for the online detection of QRS complexes in single-channel electrocardiogram (ECG) signals, which is implementable on neuromorphic devices. The proposed method incorporates continuous adaptation to the spatiotemporal dynamics of QRS morphology.
Methodology: A two-layer SNN is constructed using 10 and 50 Leaky Integrate-and-Fire (LIF) neurons with adjustable firing thresholds in the first and second layers, respectively. The streaming ECG signal is band-pass filtered within a frequency range of 8-22 Hz. Then, at each time step T, the preceding two-second chunk of the amplitude-squared signal is encoded using a Poisson rate coding scheme of length 10. Subsequently, these spike trains are fed to the network sample by sample. After each forward pass, Spike-timing-dependent plasticity (STDP) learning rule, augmented with a novel adaptive homeostatic synaptic plasticity and excitability modulation mechanism, is applied to update inter-layer connection weights. The self-regulating learning mechanism involves adjusting the weights and membrane potential thresholds of the neurons based on the timing of spikes and their firing rates deviations to stabilize the network activity in accordance with QRS complexes in each ECG segment. R-peak localization is achieved by detecting local maxima of the amplitude-squared signal during periods of burst spiking activity.
Results: Evaluation on six publicly accessible ECG datasets (MIT-BIH, QTDB, SVDB, TWADB, NSRDB, and NSTDB) demonstrated a mean sensitivity of 97.28% and a detection error rate (DER) of 3.25% in single-channel, online detection across over 2.2 million QRS complexes.
Conclusion: Comparative analysis against state-of-the-art QRS detection algorithms indicates the superiority of the proposed method and its robustness for real-time implementation.