Lead Reconstruction Using Artificial Neural Networks for Ambulatory ECG Acquisition

Figure 1. ADAS1000 based board for high quality ECG signal recording.

Figure 2. Reconstruction methodology.

Figure 3. Methodology of P wave segmentation.

Figure 4. Block diagram of the P wave segmentation algorithm to reconstruct the ECG signal.

Figure 5. ANN to reconstruct ECG: (a) single lead of the Standard 12-Lead system and (b) all leads of the Standard 12-Lead System simultaneously.

Figure 6. Additional electrodes location. In blue dots, the Dower location, and in red and purple, our proposed location. Red dots (V2, P7 and P8) are located in the patient’s chest, and the purple one (B8) is placed in the back. In black, the standard precordial leads are shown for reference.

Figure 7. Boxplot graphical representation of the five FoMs for each of the four ECG reconstruction methods studied. The results for both Linear Regression methods are represented in blue and red for the simple Linear Regression method and the P-wave segmentation method, respectively. The ANN strategies are represented in green and purple for the ANN/lead and the unique ANN methods, respectively. The parameters have been divided into three graphs to facilitate their visualization and interpretation.

Figure 8. Reconstruction by the described methods of an ECG fragment of lead II of one of the patients. To perform the reconstruction, EASI leads were chosen. Two of the five FoMs, RMS error and CC, are indicated. In blue, the original signal and, superimposed in orange, the reconstructed signal.

Figure 9. Reconstruction of the Standard 12-Lead System applying the 12-output ANN-based algorithm for the set of electrodes and leads chosen as the best option. First position was obtained by electrodes V2, P7, P8 and B8; arranged in leads P8-P7, V2-B8 and P8-B8. The original 12 leads are represented in blue, and the 12 reconstructed leads are superimposed in red.

Figure 10. Results for one of the FoMs (the RMS error) where the reconstruction is evaluated over several days at different time instants. To carry out this reconstruction, the algorithm (ANN for each lead) was only trained with the first 16 s of the first record, and the model performed the reconstruction of the rest of the records, without any feedback. The x-axis shows the records, identified with the day D and the time xxHxx at which they were taken, and the y-axis shows the standard leads, as well as their mean RMS Error value for each record.