Cardiovascular diseases are the leading cause of mortality worldwide. Electrocardiogram (ECG) is a principal, non-invasive method used to diagnose heart diseases. The need for rapid and accurate automatic analysis of ECG signals based on signal processing and machine learning has long been recognized. We present a classifier of 8 cardiac abnormalities which was trained on PhysioNet/Computing in Cardiology Challenge 2020 database containing 6,877 12-lead ECG recordings. We employed convolutions with novel, specifically designed kernels for feature extraction. Detection of QRS complexes involved a thresholding adopted from the Pan-Tompkins algorithm. The properties of P, R, T waves and PR, ST segments were determined from the averages of appropriate ECG sections. In our first approach we used a single random forest to classify ECG recordings using features from all 12 ECG leads. For training data, such classifier had F1 and F2 scores equal to 0.839 and 0.832, respectively (10-fold cross-validation). For the testing data, F2 was equal to 0.765 (PWR-UM team, run 4). We developed also a second classifier made up of 12 random forests (one for each ECG lead) and a multilayer perceptron as meta-classifier. The latter approach which on the training data yielded F1=0.837 and F2=0.845 will be tested in the second phase of the competition.