A Locally Adaptive Multi-Resolution Encoding Framework for Spiking Neural Network based ECG Beat Classification

Objectives This study highlights the importance of an appropriate encoding technique in Spiking Neural Networks, and proposes the use of multiresolution encoding, which combines multiple encodings based on input features. Methods This research experiments on The MITBIH dataset was split into train-test data using AAMI standards. It uses the Pan-Tompkins method to extract features from ECG signals and feed these features to the SNN model post-encoding. Multiple encoding schemes were implemented, including rate, temporal, population-based, and multiresolution encoding. All these encodings were compared with respect to macro-recall, accuracy, and local spike characteristics, such as Spike Efficiency (SE) and Temporal Dispersion (TD). Findings These results confirm that encoding quality in SNNs should not be judged only by accuracy or recall independently, but by the joint relationship between spike efficiency, temporal dispersion, and classification performance. The proposed clinically guided Multi-Resolution encoding offers a practical and novel solution for ECG beat classification by integrating physiological knowledge directly into spike generation with the highest accuracy of 89%. Novelty This work recommends a locally adaptive multiresolution spike encoding framework based on the local morphological characteristics of the beat features for ECG beat classification using SNNs. Unlike existing works that focus mainly on accuracy for model performance, this work focuses on accuracy and macro-recall and necessitates encoding with a balanced SE and TD.