Inter-patient automated arrhythmia classification: A new approach of weight capsule and sequence to sequence combination

Objective: We propose a new capsule network to compensate for the information loss in the deep convolutional networks in previous studies, and to improve the performance of arrhythmia classification.

Methods: We proposed the innovative weight capsule model which uses a weight capsule network combined with sequence-to-sequence (Seq2Seq) modeling to classify arrhythmia, and explored the performance of this approach.

Results: Based on the MIT-BIH arrhythmia database, we obtained better results compared with previous studies without data enhancement and balance for the samples. The specific performance was as follows: accuracy (ACC) = 99.85%; Class N: sensitivity (SEN) = 99.66%, positive predictive value (PPV) = 99.97%, specificity (SPEC) = 99.72%; Class S: SEN = 99.56%, PPV = 92.23%, SPEC = 99.68%; Class V: SEN = 99.97%, PPV = 99.38%, PPV = 99.96%; Class F: SEN = 93.81%, PPV = 100.00%, SPEC = 100.00%. When only half of the training sample was used, the method showed that the average accuracy and sensitivity of Class V and F were 1.57%, 2.01%, and 1.55% higher, respectively, than the traditional machine learning algorithm using the whole training sample.

Conclusion: Applying a weight capsule network combined with a Seq2Seq model in the field of arrhythmia not only alleviates the problem of inter-category sample imbalance effectively, but also improves the arrhythmia classification.

Significance: Our study suggests a new idea for solving the problem of small sample sizes and inter-category sample imbalance in the medical field.