Exploring a new frontier in cardiac diagnosis: ECG analysis enhanced by machine learning and parametric quartic spline modeling
Alka Mishra,Surekha Bhusnur,Santosh Kumar Mishra,Pushpendra Singh
DOI: https://doi.org/10.1016/j.jelectrocard.2024.05.086
Abstract:The heart's study holds paramount importance in human physiology, driving valuable research in cardiovascular health. However, assessing Electrocardiogram (ECG) analysis techniques poses challenges due to noise and artifacts in authentic recordings. The advent of machine learning systems for automated diagnosis has heightened the demand for extensive data, yet accessing medical data is hindered by privacy concerns. Consequently, generating artificial ECG signals faithful to real ones is a formidable task in biomedical signal processing. This paper introduces a method for ECG signal modeling using parametric quartic splines and generating a new dataset based on the modeled signals. Additionally, it explores ECG classification using three machine learning techniques facilitated by Orange software, addressing both normal and abnormal sinus rhythms. The classification enables early detection and prediction of heart-related ailments, facilitating timely clinical interventions and improving patient outcomes. The assessment of synthetic signal quality is conducted through power spectrum analysis and cross-correlation analysis, power spectrum analysis of both real and synthetic ECG waves provides a quantitative assessment of their frequency content, aiding in the validation and evaluation of synthetic ECG signal generation techniques. Cross-correlation analysis revealing a robust correlation coefficient of 0.974 and precise alignment with a negligible time lag of 0.000 s between the synthetic and real ECG signals. Overall, the adoption of quartic spline interpolation in ECG modeling enhances the precision, smoothness, and fidelity of signal representation, thereby improving the effectiveness of diagnostic and analytical tasks in cardiology. Three prominent machine learning algorithms, namely Decision Tree, Logistic Regression, and Gradient Boosting, effectively classify the modeled ECG signals with classification accuracies of 0.98620, 0.98965, and 0.99137, respectively. Notably, all models exhibit robust performance, characterized by high AUC values and classification accuracy. While Gradient Boosting and Logistic Regression demonstrate marginally superior performance compared to the Decision Tree model across most metrics, all models showcase commendable efficacy in ECG signal classification. The study underscores the significance of accurate ECG modeling in health sciences and biomedical technology, offering enhanced accuracy and flexibility for improved cardiovascular health understanding and diagnostic tools.