Can an LS-SVM classifier using RR variability and ECG spectrum measures accurately detect atrial fibrillation and other rhythms from short, single-lead ECG recordings?
An LS-SVM classifier using RR variability and ECG spectrum features shows promising performance for detecting atrial fibrillation from short, single-lead ECGs.
Aims: Atrial fibrillation (AF) is one of the principal cause of mortality in elderly, thus its detection is extremely clinically relevant. The aim of this study was to classify short, single lead, ECG recordings, as atrial fibrillation, normal sinus rhythm, other type of rhythms or noisy signal. Methods: First, we extracted, both from the ECG signals and from the RR interval series, about fifty features characterizing these four classes. Then, we applied the stepwise linear discriminant analysis for dimensionality reduction selecting a subset of thirty discriminating features. A Least Squares Support Vector Machine (LS-SVM) classifier using these features was tuned and trained on the dataset of the Physionet/Computing in Cardiology Challenge 2017. Results: The LS-SVM classifier provided, on the hidden test set of the Challenge, an official final score F1= 0.81, obtaining the twelfth place in the ranking of results with only 2 cents from the best (0.83). Conclusions: This approach seems promising in particular in detecting atrial fibrillation. Further work is needed to improve the discrimination of other rhythms and noisy signals.
Billeci et al. (Thu,) studied this question.