Multi-Feature Fusion and Compressed Bi-LSTM for Memory-Efficient Heartbeat Classification on Wearable Devices

TL;DR

This work tackles memory-efficient ECG heartbeat classification for wearable devices by integrating a multi-feature fusion strategy with Bi-LSTM networks. It introduces a framework that fuses six time-interval features $T_{ij}$ and four under-the-curve areas $A_{ij}$, detected via robust R,P,T,Q,S point extraction, and feeds these into Bi-LSTM classifiers. The approach achieves high class-specific accuracy across N, RBBB, LBBB, PVC, and PB with small to medium model sizes, and demonstrates substantial memory reductions through post-training quantization (FP16, INT8, DRQ) without sacrificing performance, enabling deployment on low-power MCUs. The results show significant improvements for challenging RBBB and LBBB classes and establish a practical path toward wearable-ready, memory-efficient ECG classification with competitive or superior performance relative to state-of-the-art methods. This has meaningful implications for accessible, real-time cardiovascular monitoring in resource-constrained settings and broadens the feasibility of continuous ECG analytics on wearables.

Abstract

In this article, we present a resource-efficient approach for electrocardiogram (ECG) based heartbeat classification using multi-feature fusion and bidirectional long short-term memory (Bi-LSTM). The dataset comprises five original classes from the MIT-BIH Arrhythmia Database: Normal (N), Left Bundle Branch Block (LBBB), Right Bundle Branch Block (RBBB), Premature Ventricular Contraction (PVC), and Paced Beat (PB). Preprocessing methods including the discrete wavelet transform and dual moving average windows are used to reduce noise and artifacts in the raw ECG signal, and extract the main points (PQRST) of the ECG waveform. Multi-feature fusion is achieved by utilizing time intervals and the proposed under-the-curve areas, which are inherently robust against noise, as input features. Simulations demonstrated that incorporating under-the-curve area features improved the classification accuracy for the challenging RBBB and LBBB classes from 31.4\% to 84.3\% for RBBB, and from 69.6\% to 87.0\% for LBBB. Using a Bi-LSTM network, rather than a conventional LSTM network, resulted in higher accuracy (33.8\% vs 21.8\%) with a 28\% reduction in required network parameters for the RBBB class. Multiple neural network models with varying parameter sizes, including tiny (84k), small (150k), medium (478k), and large (1.25M) models, are developed to achieve high accuracy \textit{across all classes}, a more crucial and challenging goal than overall classification accuracy.

Figures (11)

• Figure 1: Distribution of data annotations in the MIT-BIH Arrhythmia Database.
• Figure 2: ECG signal, peak and wave moving averages, and block of interest (BOI) for detection of R peaks.
• Figure 3: A sample of detected main points (PQRST) in the ECG signal.
• Figure 4: (a) Neural network architecture using LSTM layers and the six time interval features, (b) classification results for $n_1 = n_2 = 64$ and $p=0.5$, (c) classification results for $n_1 = n_2 = 128$ and $p=0.5$, (d) classification results for $n_1 = n_2 = 64$ and $p=0.25$.
• Figure 5: (a) Neural network architecture using Bi-LSTM layers and the six time interval features, (b) classification results for $n_1 = n_2 = 32$ and $p=0.5$, (c) classification results for $n_1 = n_2 = 64$ and $p=0.5$, (d) classification results for $n_1 = n_2 = 32$ and $p=0.25$.