This work intends to devise an efficient feature extraction scheme for identifying common cardiac abnormalities using the Fourier-Bessel (FB) expansion of RR-intervals. The Bessel basis, when used for representing the RR-intervals, meaningfully enhances the pathologically induced low-frequency changes in terms of FB coefficients. In addition to RR-intervals based features, morphological features are also extracted to ensure the characterization of diverse pathological variability present in the ECG signals. The former uses the application of FB expansion on RR-intervals. In the latter case, morphology-based features are computed from pre-processed ECG signals themselves. In all, 47 features are computed for each selected ECG leads and subsequently they are used to construct nine ensemble systems for classifying the ECG recordings into nine categories. In order to get optimal multi-class output from combination of nine different models, a Genetic algorithm (GA) based approach is applied to adaptively select the prediction thresholds such that to enhance the overall classification performance in terms of the Geometric mean score. While designing a model for each class, 5-fold cross-validation is applied to 80% of the training data to obtain reliable hyperparameters of the models. Then, GA is applied to the output of obtained models for the remaining 10% of the training data to get optimal thresholds. The last 10% of the training set is used for testing the reliability and robustness of the work. When evaluated with the 2020 PhysioNet/CinC Challenge dataset, the experimental outcomes demonstrate the F2 score, G2 score, and Geometric mean score of 67.9%, 41.6%, and 53.1% respectively on the hidden test data.