The paper identifies the effects of key design elements of bag-of-words models on the classification accuracy of ECG time series. Combinations of distinct encoding procedures, pooling methods, and classification strategies are tested in order to find best scenarios under which performances may be optimized. Extensive experiments conducted on real ECG recordings collected on chest and fingers indicate that sparse representations yield state-of-the-art results, and show robustness against data representation type, signal length, and codebook dimension.