Autoencoder-Based Extrasystole Detection and Modification of RRI Data for Precise Heart Rate Variability Analysis

Figure 1. Example of an ECG with a PVC. The RRIs both before and after are altered in most PVC occurrences: RRIs before the PVC become short, and the RRIs after the PVC become long to compensate for the heartbeat rhythm. This phenomenon is called a compensatory pause.

Figure 2. Influence of PVC on HRV features: SDNN (left) and LF/HF (right): The blue and red lines show the features extracted from the RRI data without PVC and with PVC, respectively. The vertical dashed line denotes the point of the PVC occurrence.

Figure 3. Example of an ECG with PAC. As the sinus rhythm resets after the PAC occurrence, only one RRI before PAC becomes short. That is, PAC does not have a compensatory pause.

Figure 4. Structure of an AE. An AE consists of an input layer, a hidden layer, and an output layer. x1,⋯,xM and x^1,⋯,x^M are the input variables and output variables of the AE, respectively. Circles denote units that express activation functions.

Figure 5. Flowchart of AED. This procedure can classify the RRI data into normal sinus, PVC, PAC, R wave detection error, or another type of arrhythmia.

Figure 6. Artificial noise for generating PVC (left) and PAC (right), which was added at random points in the RRI data. The peak height of H was randomly set as 100 ms <H<370 in this study.

Figure 7. The RRI modification result of PVC (top) and its enlarged display (bottom). The blue, red, and green lines represent the original RRI data, the RRI data with artificial PVC data, and the RRI data modified by the proposed DAEM, respectively. The modified data (the blue line) overlapped almost entirely with the original data (the green line) due to the high modification performance.

Figure 8. The HRV features extracted from the RRI data with and without PVC modification. The blue, red, and green lines represent the original HRV data, the HRV data extracted from the RRI with artificial PVC data, and the HRV data extracted from the RRI data modified by the proposed DAEM, respectively. The modified data (the blue line) overlapped almost entirely with the original data (the green line).

Figure 9. The RRI modification result of PAC. The blue, red, and green lines represent the original RRI data, the RRI data with artificial PVC data, and the RRI data modified by DAEM, respectively. This figure shows that the modified RRI was very close to the original RRI data.

Figure 10. The HRV features extracted from the RRI data with and without PAC modification. The blue, red, and green lines represent the original HRV data, the HRV data extracted from the RRI with artificial PAC data, and the HRV data extracted from the RRI data modified by the proposed DAEM, respectively. This figure shows that the HRV data extracted from the modified RRI became close to the HRV data extracted from the original RRI data.

Figure 11. The RMSE decrement rates of the RRI data with PVC by DAE, PLS, and LWPLS. The proposed DAE achieved the best performance among the three methods.

Figure 12. The RRI errors in the R wave detection error modification. Although the average RRI improvements were almost the same, the deviation of DAE was smaller than that of the other two.