Recent progress in computerized electrocardiogram (ECG) interpretation makes such systems a useful tool in diagnostics of cardiovascular diseases. However, questions remain on how well these systems can work in situations of reduced-lead data. The ability to successfully diagnose diseases based such data would help in adoption of lower-cost, reduced-lead devices. This is the topic of the PhysioNet/Computing in Cardiology Challenge 2021, which asks the participants to train machine learning models for interpretation of twelve-lead, six-lead, three-lead, and two-lead ECGs, and use the models to identify one of 27 cardiac conditions, such as atrial fibrillation, bradycardia, sinus arrhythmia, etc. The training data provided by the competition’s organizers are 43,101 ECGs from five different sources in several countries. The organizers use a private data set of 16,630 ECGs to test the model’s accuracy. To address this Challenge, our team (team northern_lights) built a 1-dimensional Convolutional Neural Network (CNN) with residual connections. To capture long-range correlations, we selected a large kernel size spanning a complete R-R interval and used a large number of kernels (100 to 200). Patients’ age and sex were also added as an input to the model, along with extracted heart rate. To avoid overfitting, we used a data augmentation technique, randomly shifting each ECG during the training process. We were able to achieve the challenge score of 0.78 (12-lead) on the training data and the Challenge test data results 0.413, 0.364, 0.381, and 0.366 on 12-lead, 6-lead, 3-lead, 2-lead validation datasets respectively. The results show that the reduced-lead ECG can be used for diagnostics, albeit with somewhat lower accuracy.