Detection of paroxismal atrial fibrilation with normalized heart rate variability analysis

Paroxysmal Atrial Fibrillation (PAF) is the initial stage of Atrial Fibrillation, one of the most common arrhythmia types. Although PAF is not directly fatal, it triggers other deadly conditions and increases the risk of stroke. Moreover, during the PAF episode, the patient may be in an action that affects his and others' lives. Therefore, both the early diagnosis to start treatment immediately and the prediction of attacks to warn the patient to take a safe place are essential. In this study, traditional machine learning methods including deep learning were evaluated for both tasks. In this thesis, an open-access database that consists of 30-minute ECG data from 49 healthy individuals, 24 PAF patients with no recent attack, and 25 PAF patients with the recent attack was used. Heart Rate Variability (HRV) data and heart rate normalized HRV (NHRV) data were obtained from these ECGs. Commonly used time-domain, frequency-domain, wavelet transform, and nonlinear features were extracted for both data. Feature normalization methods were also utilized for these features. The genetic algorithm chose the more valuable features. Extracted and selected features are applied to inputs of k-nearest neighbors, multi-layer perceptron, support vector machines with radial basis kernel function, and convolutional neural-network-based deep learning classifiers. As a result, both tasks resulted in 100% classifier performances with the deep learning algorithm using NHRV features with z-score normalization. Moreover, although deep learning runs with raw data in the literature, feature extraction may be efficient in problems that require specific feature extraction methods such as HRV analysis.