Combining symmetric projection attractor reconstruction with machine learning to automatically detect atrial fibrillation during hemodialysis

Abstract

Atrial fibrillation (AF) is highly prevalent among patients suffering from renal failure and linked to a doubled one-year mortality rate in this group. Implementing a reliable algorithm that can detect AF on Electrocardiograms (ECGs) recorded during patients’ regular dialysis treatments, would allow for routine monitoring, without further increasing patients’ hospital time. Unfortunately, ECGs recorded during dialysis are harder to interpret via automated algorithms, because the ECG waveform is altered by fluid and electrolyte shifts. Symmetric projection attractor reconstruction (SPAR) is a new method of analyzing cardiovascular waveform data, that may be able to overcome these challenges. SPAR uses delay coordinates to convert a short time series signal into a so-called attractor, bound in three- or higher-dimensional phase space. Its rotationally symmetric, two-dimensional projections emphasize different aspects of the original signal. This can be used to better understand the underlying waveform. This thesis combines SPAR with stacked k-nearest neighbor classifiers to automatically detect AF from a single lead ECG. The methodology was first established on ECGs from the open access database PhysioNet and then applied to ECGs recorded during dialysis. 30s ECG samples were taken from three different PhysioNet databases and extracted from 24h ECGs recorded during a study on end stage renal disease, at six different timepoints during and after dialysis. Models were trained on six different training and test compositions, each containing an approximately equal number of AF and control records. Controls for the dialysis records were matched based on age, gender and dialysis vintage in months. SPAR was used to generate 20 attractor projections from every ECG sample. These were quantified by calculating their angular density distribution, radial density distribution and attractor outline. By training one k-nearest neighbor classifier on each of the density curves, stacked models of 60 classifiers were trained for each training set. Test records were classified based on the mean posterior probability predictions of all classifiers in the model, that have a cross-validated accuracy of 70% or more based on 10-fold cross validation. All algorithms and data processing steps were implemented in MATLAB®. The highest performing PhysioNet model achieved classification accuracies of 89.3% and 93.8% on the two PhysioNet test sets. The model trained on samples at the beginning, middle and end of dialysis classified samples from the start of dialysis test set with and average accuracy of 85.7%. Including both PhysioNet and dialysis samples in the training set showed no improvements for the classification of either category. The visual comparison of PhysioNet and dialysis three-point attractors and their densities also showed that the differences between the two groups exceed the differences between AF and no AF records of either group. This confirms that changes in the ECG caused by the dialysis are also visible in the ECG attractors. These classification accuracies achieved by the models in this thesis compare well to other attempts at automated AF detection in the literature. Most of these approaches rely on convolutional neural networks or meticulously selected features. Compared to these attempts, the methodology presented in this thesis has the advantage of a simpler, less computationally expensive methodology, good stability towards outliers and noise and no necessity for feature selection. Its ability to classify very short ECG samples in a short time qualifies it for real time monitoring applications. Experimenting with even less preprocessing and investigating methods of feature selection may be ways to further improve AF detection using SPAR.