Key result
A neural network classifier achieved the highest accuracy for multi-type arrhythmia classification with an F1 score of 0.72, outperforming linear discriminant analysis (F1=0.66) and classification trees (F1=0.65).
Why the study?
Reliable AF detection is hindered by its episodic appearance, diverse arrhythmias with irregular RR intervals, and ECG noise, prompting methods to classify short, single-lead ECG signals.
Absolute Event Rate: 0.72% vs 0.66%
A neural network approach using 118 time and frequency domain features achieved the highest accuracy (F1=0.72) for classifying short, single-lead ECGs into various rhythms.
Atrial fibrillation (AF) is associated with significant risk of heart failure and consequent death. Its episodic appearance, the wide variety of arrhythmias exhibiting irregular AF-like RR intervals and noises accompanying the ECG acquisition, impede the reliable AF detection. Therefore, the Computing in Cardiology Challenge 2017 organizers encourage the development of methods for classification of short, single-lead ECG as AF, normal sinus rhythm (NSR), other rhythm (OR), or noisy signal (NOISE). This study presents a set of 118 time and frequency domain feature including descriptors of the RR and PP intervals; QRS and P-wave amplitudes; ECG behavior within the TQ intervals, deviation of the TQ and PQRST segments from their first principle component analysis vector; dominant frequency; regularity index, width and area of the power spectrum estimated for the ECG signal with eliminated QRS complexes. Three classification techniques have been applied over the 118 ECG features – linear discriminant analysis (LDA), classification tree (CT) and neural network (NN) approach. The scores over a test subset are: (i) FNSR = 0.81; FAF = 0.61; FOR = 0.53, F1 = 0.65 for CT, which is the most simple model; (ii) FNSR = 0.82; FAF = 0.62; FOR = 0.53, F1 = 0.66 for LDA, which is the model with the most reproducible accuracy results; (iii) FNSR = 0.86; FAF = 0.74; FOR = 0.57, F1 = 0.72 for NN, which is the most accurate model.
No takes yet. Share an insight, caveat, or question.
Jekova et al. (2020) studied Arrhythmia (Atrial Fibrillation) (n=12,186). Neural Network classifier vs. Classification Tree and Linear Discriminant Analysis was evaluated on F1 score for multi-type arrhythmia classification on a test subset. A neural network classifier achieved the highest accuracy for multi-type arrhythmia classification with an F1 score of 0.72, outperforming linear discriminant analysis (F1=0.66) and classification trees (F1=0.65).
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: