Reconstructing 12-lead ECG from reduced lead sets using an encoder–decoder convolutional neural network

Fig. 1.. Different components of an ECG waveform. This heartbeat is sampled from a healthy control patient (P260) from the PTB database.

Fig. 2.. Preprocessing flow. Following the preprocessing pipeline, each lead recording of a heartbeat is represented by the real and imaginary components of its short-time Fourier transform.

Fig. 3.. Inputs and their maps. As shown above, initially each heartbeat is isolated. Then, for each heartbeat, different pairs of data (input and label) are derived based on different available independent leads as inputs. We provide an input lead map along with every three-lead input, showing which leads are present. The Final goal is to generate the 12-lead ECG as output.

Fig. 4.. (a) Network architecture. This network is a modified convolutional encoder–decoder setup. N represents the number of input channels and is 7 for 3-lead reconstruction with input map, 6 without the input map, and 2 for 1-lead reconstruction scenario. The values listed under each block represent the dimensions of that block. (b) An example of the proposed model’s input and output for 3-lead reconstruction. This heartbeat instance is randomly chosen from PTB database. The STFT of three input leads (in this example, leads I,II, and V1) are fed to the proposed model to reconstruct the STFT of 12-lead ECG. Next, the inverse STFT is applied to get the signals in time domain. A colormap is displayed in the top right corner.

Fig. 5.. ECG reconstruction results with and without input lead maps. (a) Average CC for each patient. Note that CC is a value between −1 and 1, where closer values to one represent more similarities between the original and reconstructed signals. These results indicate that both scenarios achieve similar outcomes. (b) Boxplot graphical representation of correlation coefficients of different leads for each patient. (c) Average RMSE for each patient. Please note that RMSE values are calculated from normalized signals.

Fig. 6.. ECG reconstruction sample. This figure shows a reconstructed heartbeat for patients 1 (on the left) and 3 (on the middle) from the BCM dataset and patient 269 from PTB database. The x axis is the time sample and the y axis is the normalized voltage of the waveform. The blue trace shows the original ECG of the patient, and the orange trace shows the reconstructed ECG using the CNN model. The model receives three leads as input, and reconstructs the rest. Each input includes a map represented as an extra channel.

Fig. 7.. Average MI values. This figure shows the average MI between each lead and other eleven leads for each patient.

Fig. 8.. Conditional Entropy. (a) These colormaps show the conditional entropy values for three patients. Considering li(Pk) to stand for lead li of patients Pk, the color of the cell located at row i and column j, represents H[li(Pk)∣lj(Pk)] which is a measure of how much knowing lead lj of patients Pk reduces the uncertainty about lead li of the same patient. Here, we show the results for patients P1,P9, and P15. The lower the conditional entropy, the more information we have on li when knowing lj. As expected, the diagonal elements are zero since H[li(Pk)∣li(Pk)]=0. (b) The colormap shows 112ΣnH[ln(Pj)∣li(Pj)],ln,li∈l0,l1,…,l11. Looking at different patients’ data shows that for most cases, knowing leads I,II,V4:V6 furthest reduces the uncertainty of other leads. (c) This plot shows the row-wise average of the colormap in b. (d) For every patient, entropy of one of it’s leads conditioned on the same lead of patient 1 is shown here. As expected, entropy values indicate that different patients have less information in common, and reconstructing a new patient’s ECG signals based on another patient’s is a harder task than patient-specific reconstruction. (e) A more general representation…

Fig. 9.. MI and CC relation. For each patient, average MI, minimum MI, and average reconstruction CC are shown (in blue, gray, and orange, respectively). Here, ρ‾li,li′(P2) represents the average CC between the original and reconstructed lead li over testing data for patient P2 from the BCM dataset. Note that the reconstruction results belong to patient-specific models which accept any ECG lead as input. According to these plots, a direct relation between higher MI and better reconstruction is observed.

Fig. 10.. Entropy and CC relation. For each lead, average entropy and reconstruction CC over the BCM dataset’s patients are shown (in blue and maroon, respectively). Please note that ρ‾Ili,li′(Pk) represents the average reconstruction correlation coefficient for patient Pk when lead I is used as input. As expected, an inverse relationship between these values can be observed.