An optimization method based on genetic algorithm for heart rate variability analysis in the prediction of the onset of cardiac arrhythmia

Abstract

Heart rate variability (HRV) is one of the common biological markers for developing a diagnostic system of cardiovascular disease. HRV analysis is used to extract statistical, geometrical, spectral and non-linear features in such diagnostic system. The diagnostic accuracy can be maximized by applying a feature selection step that selects an optimal feature subset from the extracted features. However, there are shortcomings in using only the feature selection for optimizing a diagnostic system that is based on HRV analysis. One of the main limitations is that the parameters of HRV feature extraction algorithms are not optimized for maximal performance. In addition, the feature selection process does not consider the feature cost and misclassification error of the selected optimal feature subset. Therefore, this thesis proposes a multi-objective optimization method that is based on the non-dominated sorting genetic algorithm to overcome these shortcomings in a cardiac arrhythmia prediction system. It optimizes the HRV feature extraction parameters, selects the best feature subset, and tunes the classifier parameters simultaneously for maximum prediction performance. The proposed optimization algorithm is applied in two cardiac arrhythmia cases, namely the prediction of the onsets of paroxysmal atrial fibrillation (PAF) and ventricular tachyarrhythmia (VTA). In the proposed approach, trade-off between multiple optimization objectives that contradict to each other are also analyzed. The optimization objectives include the feature count, measurement cost, prediction sensitivity, specificity and accuracy rate. The following results prove the effectiveness of the proposed optimization algorithm in the two arrhythmia cases. Firstly, the PAF onset prediction achieves an accuracy rate of 89.6%, which significantly outperforms most of the previous works. This accuracy rate is achieved even with the HRV signal length being reduced from the typical 30 minutes to just 5 minutes (a reduction of 83%). In the case of VTA onset prediction, the accuracy rate of 78.15% is achieved with 5-minute signal length. This result outperforms previous works. Another significant result is the sensitivity rate improvement with the tradeoff of lower specificity and accuracy rate for both PAF and VTA onset predictions. For instance, the sensitivity rate of the VTA onset prediction system improved from 81.48% to 92.59% while the accuracy rate reduced from 78.15% to 72.59%.