Machine Learning Approaches for Heart Rate Variability Data Correction and Coronary Artery Calcification Classification

Abstract

Coronary artery disease (CAD) is one of the most common cardiovascular diseases and a major cause of death worldwide. Detection of coronary artery calcification (CAC) through coronary computed tomography angiography (CCTA) is normally used to diagnose CAD. Previous studies have demonstrated significant differences in the physiological response to exercise between individuals with and without CAC. This thesis aimed to apply machine learning (ML) methods on data measured during exercise to predict the presence of CAC, with a particular focus on the analysis of heart rate (HR) and heart rate variability (HRV) measured with HR chest straps. For this purpose, signal issues from the HR monitors must be handled appropriately. Hemodynamic measures were collected from healthy participants before, during, and after a 91-km mountain bike race (the North Sea Race). The presence of CAC was determined by CCTA after the race. Several time series methods were applied to the HRV data to address data artifact correction. A statistical analysis of hemodynamic measures at the most challenging hill was conducted to determine physiological differences between individuals with and without CAC. Finally, various classification algorithms were used to predict the presence of CAC based on hemodynamic and HRV data. In Papers 1 and 2, the autoregressive integrated moving average method was shown to outperform other artifact correction methods for HRV data, even with minimal training data and computational cost. Cubic interpolation, the most common artifact correction method, was found to be less effective and is therefore not recommended. Paper 3 demonstrated that during prolonged high-intensity endurance exercise, diastolic blood pressure and HRV were the most important predictors of the presence of CAC. The level of physiological strain seems to be an essential factor in inducing these differences in otherwise healthy individuals. In Paper 4, an ML approach combining dimensionality reduction with logistic regression achieved 84% accuracy for classifying individuals with and without CAC. This model’s most important input features were blood pressure, age, HRV, and body mass index. Overall, the results suggest that feature-based statistical analysis of HR and HRV data is more effective than raw-signal analysis, likely due to the high volatility of the signal data.