Patient-Specific Deep Architectural Model for ECG Classification

Abstract

Heartbeat classification is a crucial step for arrhythmia diagnosis during electrocardiographic (ECG) analysis. The new scenario of wireless body sensor network- (WBSN-) enabled ECG monitoring puts forward a higher-level demand for this traditional ECG analysis task. Previously reported methods mainly addressed this requirement with the applications of a shallow structured classifier and expert-designed features. In this study, modified frequency slice wavelet transform (MFSWT) was firstly employed to produce the time-frequency image for heartbeat signal. Then the deep learning (DL) method was performed for the heartbeat classification. Here, we proposed a novel model incorporating automatic feature abstraction and a deep neural network (DNN) classifier. Features were automatically abstracted by the stacked denoising auto-encoder (SDA) from the transferred time-frequency image. DNN classifier was constructed by an encoder layer of SDA and a softmax layer. In addition, a deterministic patient-specific heartbeat classifier was achieved by fine-tuning on heartbeat samples, which included a small subset of individual samples. The performance of the proposed model was evaluated on the MIT-BIH arrhythmia database. Results showed that an overall accuracy of 97.5% was achieved using the proposed model, confirming that the proposed DNN model is a powerful tool for heartbeat pattern recognition.

Figure 1

Schematic illustration of two heartbeat classification frameworks: (a) traditional framework and (b) the framework of proposed DL architectural model with time-frequency representation.

Figure 2

ECG preprocessing, sample heartbeat waveform from number 201 record lead II, including normal (N), premature ventricular contraction (PVC), and atrial premature contraction (APC) AAMI heartbeat classes. (a) Baseline drift elimination; (b) QRS-complex detection; (c) MFSWT spectrogram corresponding to the waveform in (b); (d)~(i) MFSWT spectrograms corresponding to each segmented heartbeat by 700 ms windows in (b). In the absence of other special instructions, all spectrograms were normalized to [0, 1].

Figure 3

Heartbeat signals and corresponded spectrograms: (a) original heartbeat and reconstructed heartbeat; (b) MFSWT; (c) CWT @ mexh wavelet (d) WVD; and (e) FSWT @ κ = 4. All spectrograms were normalized to [0, 1].

Figure 4

Stochastic corruption process in SDA model training. (a) Original spectrogram. (b) Result of stochastic corruption process (P = 0.5).

Figure 5

Schematic view of SDA for feature learning and DNN-based classification: (a) three layers of SDA and four layers of DNN; (b) part of the first layer weights of unsupervised trained SDA model; (c) extracted features of number 100 record first heartbeat; and (d) details of the weights marked with a red box.

Figure 6

The workflow of patient-specific DNN training.

Figure 7

Heartbeat classification results by patient-specific models with different parameters. The best results of SVEB and VEB classification from other works in Table 4 are shown as benchmarks; the best results of the proposed models are marked by asterisks. A~G represent the sizes of the first hidden layer which are 64, 128, 256, 512, 1024, 2048, and 4096, and the number of neurons of the next layer was set as half of the current layer if SDA is a multilayer structure.

Figure 8

Relationships between classification performance and the number of individual samples were used for model training.