Machine Learning-Based Classification of Cardiac Diseases Using Spectral and Non-linear Analysis of Electrocardiogram Signals

Abstract

Cardiovascular Diseases (CVDs) are one of the leading health challenges worldwide. According to the World Health Organization (WHO) report, the latest global health estimates indicate that CVDs account for approximately 32% of all deaths globally, resulting in around 17.9 million fatalities annually. Among these deaths, Myocardial Infarction (MI), Coronary Artery Disease (CAD), and Congestive Heart Failure (CHF) present particularly severe outcomes, often occurring without warning and serving as leading causes of death. Most CVDs can be alleviated or prevented if detected and treated in their early stages. Therefore, advancing investigative methods such as spectral and nonlinear, particularly those driven by machine learning, could significantly enhance individual health. Detecting R peaks within the QRS complex of the Electrocardiogram (ECG) signal is pivotal for accurately diagnosing heart conditions. However, denoising ECG signals and accurately detecting R peaks present significant challenges due to various noise factors such as baseline wander, motion artifacts, muscle contraction noise and power line interference. To address these challenges, this study proposed an efficient R peak detection technique that leverages the Maximal Overlap Discrete Wavelet Transform (MODWT) in combination with the FejerKorovkin (FK) wavelet function. The simulation result has shown that the proposed method achieved a sensitivity (Se) of 99.97%, a Positive Prediction (PP) rate of 99.64%, a low Detection Error Rate (DER) of 0.04 for T-Wave Alternans Database (TWADB) and a Se of 99.95%, a PP rate of 99.99%, and an extremely low DER of 0.063 for QT Database (QTDB). After R-peak detection, this study has been extended to Heart Rate Variability (HRV) signal analysis using spectral methods. The non-stationary nature of HRV signals necessitates the adoption of time-frequency methodologies that can capture and portray the temporal changes in frequency components. Further investigating the intricate relationship between cardiovascular signals, including HRV, respiratory (RESP), Systolic Arterial Blood Pressure Variability (SABPV), and Arterial Pressure Interval Variability (APIV) of healthy Young and Elderly subjects to understand cardiovascular control. This study proposed an Optimized Gaussian Window Reassignment(OptGWRS)TimeFrequency(TF)techniquetocapturethecharacteristics of these signals. The parameters of the Gaussian window are optimized using a Particle Swarm Optimization (PSO) algorithm to maximize the energy concentration of signals in the TF domain. The simulation results have shown that Non-Stationary Spectral Coherence (NSSC) ranges between the HRVandAPIVsignalsinthelowfrequencybandhavebeensignificantlyreduced(p=0.000001) in the elderly group when compared to the young group. In the LF band, the NSSC among HRV and RESP signals have not been influenced by ageing and in the HF band, it is significantly reduced in the elderly compared to the young group (p =0.0125). For Myocardial Infarctions (MIs) detection, the low amplitude and short duration of ECG signals pose diagnostic challenges. To address these challenges, this study proposed an automated multi-classification model integrating Optimized Modified Stockwell Transform (OptMST), Multiscale Entropies (MSEns) and Machine Learning (ML) techniques, such as Naive Bayes (NB), Support Vector Machine (SVM), Extreme Learning Machine(ELM),AdaptiveK-Nearest Neigh bors (A-KNN), Bayesian Optimization SVM (BO-SVM), and Random Forest (RF) to classify ten MI types and healthy controls. Initially, the ECG signal is subjected to OptMST up to four levels of frequency band decomposition. Subsequently, five non-linear features, such as Multiscale Approximate entropy (MS-ApEn), Multi-scale Sample entropy (MS-SampEn), Multi-scale Entropy of Entropy (MS-EnoEn), Multi-scale Increment Entropy (MS-IncEn), and Multi-scale Phase Entropy (MS PhasEn) are extracted in each frequency band of OptMST coefficients. After this, the ReliefF ranking algorithm is employed to select the top ten features out of twenty, aiming to enhance predictive performance, improve computational efficiency, and reduce the risk of over-fitting of ML. The simulation results demonstrated that the highest classification performance has been achieved by the RF model with an accuracy (Ac) of 98.44%, a recall of 94.6%, a Specificity (Sp) of 98.93%, a Precision (Pr) of 94.67%, and a F-score of 94.62%. Ex panding our scope to broader cardiac conditions, a hybrid intelligence system combining Layer Fusion Convolutional Neural Network (LF-CNN) with MLtechniques (BO-SVM, KNN,RF) has been proposed to classify CAD and CHF subjects. Time-series and spectral features from HRV data of CAD and CHF subjects have been utilised, with batch normalisation and t-distributed Stochastic Neighbor Embedding (t-SNE) for feature fusion and visualisation. The simulation results demonstrated that the hybrid model achieved an Ac of 99.88%, a Pr of 99.89%, a recall of 99.68%, and an F1-score of 99.79%. This study can advance cardiovascular diagnostics by integrating signal processing with ML, from MODWT-based R peak detection to hybrid LF-CNN-ML systems. These proposed approaches can enhance early identification and classification of cardiac conditions, offering clinicians powerful tools to improve patient outcomes across the CVDs spectrum.