A novel multi-class classification method for arrhythmias using Hankel dynamic mode decomposition and long short-term memory networks

Abstract

The complex dynamic properties of ECG signals and the challenge of multi-category classification make automated diagnosis of arrhythmias difficult. In this paper, we propose a new model for the multiclassification task of arrhythmia, which combines Hankel Dynamic Modal Decomposition (HDMD) and Long Short-Term Memory Network (LSTM). HDMD is used to construct the Hankel matrix, the optimal delay parameter is selected based on 90% of the energy of the singular values, and dynamic modal features extracted from it are used as the input sequences of LSTM. The LSTM model is optimisation is performed by minimising the cross-entropy loss function, setting the maximum number of iterations to 60 and using an early stopping strategy to avoid overfitting. The model was validated on the MIT-BIH arrhythmia database, which contains 109,402 beats and is classified into five categories according to the AAMI criteria: normal beats (N), ventricular premature beats (V), fusion beats (F), atrial premature beats (S) and unclassifiable (Q). By comparing with the direct use of LSTM, the experimental results showed that the HDMD-LSTM model showed different degrees of improvement in all five classifications, especially in the classification of atrial premature beats (S) and ventricular premature beats (V), and the overall classification accuracy improved to 0.85. Future work can focus on two aspects: first, to solve the category imbalance problem, by over sampling or undersampling techniques to improve the classification ability of a few categories; second, exploring the embedding of DMD into the network structure of LSTM to further optimise the feature extraction and classification performance.