Importance of Electrode Selection and Number in Reconstructing Standard Twelve Lead Electrocardiograms

Abstract

Many clinical and consumer electrocardiogram (ECG) devices collect fewer electrodes than the standard twelve-lead ECG and either report less information or employ algorithms to reconstruct a full twelve-lead signal. We assessed the optimal electrode selection and number that minimizes redundant information collection while maximizing reconstruction accuracy. We employed a validated deep learning model to reconstruct ECG signals from 250 different patients in the PTB database. Different numbers and combinations of electrodes were removed from the ECG before reconstruction to measure the effect of electrode inclusion on reconstruction accuracy. The Left Leg (LL) electrode registered the largest drop in average reconstruction accuracy, from an R2 of 0.836 when the LL was included to 0.737 when excluded. Additionally, we conducted a correlation analysis to identify leads that behave similarly. We demonstrate that there exists a high correlation between leads I, II, aVL, aVF, V4, V5, and V6, which all occupy the bottom right quadrant in an ECG axis interpretation, and likely contain redundant information. Based on our analysis, we recommend the prioritization of electrodes RA, LA, LL, and V3 in any future lead collection devices, as they appear most important for full ECG reconstruction.

Keywords: 12-lead reconstruction; artificial intelligence; lead importance; lead placement; lead significance.

Figure 1

ECG electrode placement: The name and placement of the standard ten electrodes used to collect and calculate a standard twelve-lead ECG. These ten electrodes include the limb leads (RA—right arm, LA—left arm, RL—right leg, and LL—left leg) as well as the chest leads (V1–V6). Typically, the RL lead is used as a ground electrode.

Figure 2

Simplified AI reconstruction: Simplified example of model training in which artificial intelligence attempts to fill in missing leads. The loss function quantifies the difference or “error” between the model’s reconstruction and the actual patient signal. This quantified error informs the model about how faithfully it reconstructed the signal.

Figure 3

Combinatoric ECG Reconstruction: (A) Leads were removed from the same signal in a systematic, combinatoric process, and the model was tasked with reconstructing the ECG based on the remaining leads. The reconstructions were compared against the actual signal to calculate performance metrics. (B) The equation that governs the combinatoric inclusion and exclusion of ECG lead information, explicitly defining the number of missing lead ECGs created from just one patient ECG.

Figure 4

Lead Frequency in top 10% of Reconstructions: The top 10% of reconstructions for each of the number of lead combinations. For each of the best reconstructions, the specific leads used to reconstruct the signal were tallied and reported based on how frequently they were used. This was normalized based on how frequently each lead would be chosen if reconstruction performance was randomized (represented by 0 on the y-axis). Leads that appear more frequently than chance are shown as positive frequencies while leads chosen less frequently appear as negative.

Figure 5

Lead Cross-Correlation Matrix: The cross-correlation matrix quantifying the extent to which two leads were linearly correlated. A value of 1 denotes perfect linear correlation, −1 perfect inverse correlation, and 0 no correlation. Values close to 0 are depicted as darker blue. Values on the diagonal (when each lead was compared against itself) all equal 1. This matrix is a Hankel matrix—symmetric when reflected across the descending left-to-right diagonal.

Figure 6

Electrode Frequency in the Top and Bottom 20% of Reconstructions: The electrode frequencies shown in the top and bottom 20% of reconstructions. For each of the best reconstructions, the specific electrodes used to reconstruct the signal were tallied and reported based on how frequently they were used. This frequency was normalized based on how frequently the electrode would be chosen if reconstruction performance was randomized (represented by 0 on the y-axis). Electrodes that appeared more frequently than chance are shown as positive frequencies while those chosen less frequently appear as negative.

Figure 7

Absolute Signal-to-Noise Ratio ($\left|SNR\right|$ for each Signal: (A) A scatterplot of the absolute value of SNR for each of the 1,023,500 signals used in this analysis. (B) A box-and-whisker plot for each signal binned based on its |SNR|. Outliers are visualized as black diamonds.