Evaluation of the linear and nonlinear classifiers for distinguishing between healthy subjects and patients with valvular heart diseases based on electrocardiograms, seismocardiograms, and gyrocardiograms.

BACKGROUND AND

OBJECTIVE

Heart rate variability (HRV) is a prognostic marker in numerous cardiovascular and non-cardiovascular conditions. Valvular heart disease (VHD) is a cardiovascular disease that affects the heart valves (aortic valve, mitral valve, pulmonic valve and tricupsid valve) and is the third most common cardiovascular disease. Traditional methods, such as echocardiography, computed tomography, and magnetic resonance imaging, are effective, but their limitations in outpatient monitoring have led to the exploration of alternative techniques, such as electrocardiography (ECG), seismocardiography (SCG) and gyrocardiography (SCG). In this study, we evaluated seven methods for differentiation between healthy volunteers and patients with valvular heart diseases: three linear classifiers (Logistic Regression, Support Vector Machine with a linear kernel, Ridge Regression) and four decision tree-based models (Random Forest, Bagged Trees, Gradient Boosting, Extreme Gradient Boosting).

METHODS

The study was carried out in two publicly available data sets with concurrent electrocardiographic (ECG), seismocardiographic (SCG), and gyrocardiographic (GCG) signals (Mechanocardiograms with ECG reference and An Open-access Database for the Evaluation of Cardio-mechanical Signals from Patients with Valvular Heart Diseases) that have 29 and 30 simultaneous recordings, respectively. All classifiers were trained on HRV indices calculated from concurrent ECG, SCG, and GCG signals from both datasets. Heartbeats in the SCG and GCG signals were detected as local maxima delayed from the locations of QRS complexes in the ECG signal by maximally 150 ms.

RESULTS

The results showed that linear and tree-based classifiers that work on HRV indices derived from ECG, SCG and GCG signals (accuracy of 0.9492 in the best case, 0.7627 in the worst case) could be a useful tool to differentiate between different heart diseases.

CONCLUSIONS

The use of multimodal recordings provides more comprehensive information on the state of the cardiovascular system that, in combination with machine learning-based classifiers, could help diagnose cardiovascular conditions more efficiently.