A real-time embedded system to detect QRS-complex and arrhythmia classification using LSTM through hybridized features

Abstract

The electrocardiogram (ECG) is an extremely valuable medical examination for monitoring cardiac disorders. The QRS waves on the ECG signal are essential in diagnosing these disorders. While numerous algorithms for detecting R-peaks/QRS complexes are developed, most are focused on complex computations that need off-line execution on a PC. However, advancements in telemedicine and wearable devices require an algorithm that runs effectively on an embedded system. This paper aims to design and develop an embedded system to detect the QRS complex and arrhythmia classification based on the patient-specific ECG data. The proposed model is based on the Discrete Wavelet Transform (DWT), Delta Sigma Modulation (DSM) with local maximum/minimum point algorithm to detect R peak/QRS complex. It extracts several R peaks/QRS complex features, such as the waves peak, onset, offset, and duration between consecutive R peaks (RR interval), and uses these to improve classification accuracy. We proposed Long Short Term Memory (LSTM) neural network for arrhythmia classification. First, the ECG signal is extracted through the embedded system and used for further processes. Second, the QRS complex/R peak is detected using modulated bitstreams, threshold level through DSM and DWT, respectively. Thirdly, the extracted features are hybridized and input into an LSTM for arrhythmia classification. The MIT-BIH database was used to evaluate the algorithm’s performance, and the accuracy, positive predictivity, sensitivity, and F1 score were evaluated as performance metrics. The algorithm achieved 99.64 %, 99.15 %, 99.87 %, and 98.18 % for all four metrics, respectively. The algorithm was then executed on an embedded system, and its run time and power consumption were examined. The DSM algorithm detects QRS waves in 17.2 ms, while the DWT method detects R peak in 14.02 ms. The proposed LSTM algorithm takes 58 ms for classification. The DSM chip (MCP3008 ADC) consumes 680 nW of power at a sampling rate of 500 Hz. Additionally, the algorithm’s performance was compared to those of other widely used algorithms. The suggested approach holds considerable promise for long-term monitoring in wearable systems.