INTRODUCTION: This work investigates technics that allow for the automatic classification of normal vs abnormal heartbeats with the goal of assisting general practitioners. In fact, many different ECG waveform features have been proposed over the years as inputs to normal/abnormal heartbeat classifiers. However, there is a need for the formal comparison of the classification performances obtained when using these features, and more importantly their joint combinations, on a single common dataset. This study thus investigates the classification of heartbeats as normal or abnormal using combinations of 5 different feature extraction approaches and 2 classifiers.
METHODS: The MLII lead of the MIT-BIH database was used and the detection of heartbeats was obtained from Physionet’s algorithms. Waveform features were then extracted using 5 different promising approaches from the literature, namely: Hermite basis function expansion, cumulants from higher order statistics of both the ECG and wavelet coefficients, morphological features of a heartbeat and RR intervals. All possible combinations of these features were then used as inputs to classify normal and abnormal heartbeats. Two different supervised classifiers were used: a Multilayer Perceptron (MLP) and a Support Vector Machine (SVM) where 50 302 randomized heartbeats where used for training and 14 372 for testing.
RESULTS AND CONCLUSION: The best feature set in terms of the accuracy of classification was found to be the combination of the Hermite basis function expansion, the complete higher order statistics of the ECG waveform and the RR intervals. In fact, a classification accuracy of 94.6% was obtained with the MLP for this feature set while a near perfect accuracy of 99.1% was obtained with the SVM (normal: precision=98.7%, sensitivity=99.5%, F1=99.1%; abnormal: precision=99.5%, sensitivity=98.7%, F1=99.1%). This feature set classified with the SVM was thus found to outperform all other feature combinations studied.