If you are interested by heart evoked potentials, you are probably averaging a lot of data (several hundreds of trials). In this context, all the noise is probably going away without any pre-processing (except maybe for the high-pass filter).
You could try for one subject to reproduce the same analysis without ICA and without low-pass filter, and see if you get similar results. A general recommendation is to limit the number of pre-processing steps strictly to the minimum required for the analysis you want to perform.