The electrocardiogram (ECG) is widely used in clinical practice for diagnosing diseases. However, conventional ECGs contain 12 projections of the three-dimensional cardiac dipole vector, resulting in significant redundancy. To address this issue, we developed a metric based on mutual information to quantify redundancy in standard ECGs. We used two strategies for reducing redundancy: eliminating leads and applying linear transformations to the original ECG leads. We employed a convolutional neural network (CNN) with inputs generated from redundancy elimination. We found that reducing the input to three orthogonal three-dimensional coordinates had minimal impact on the model's performance owing its slight decrease to potential elimination of relevant information. However, using six channels, despite increased redundancy, mitigated distortion by minimizing information loss. This study establishes an objective criterion for selecting cardiac vector projections that minimize redundancy while maintaining diagnostic capacity in the CNN. This serves as an objective and robust criterion for selecting the most informative projections in wearables or Holter monitors.