The measurement of high-energy neutrinos plays an important role in the understanding of the most energetic events in the universe. Neutrinos with an energy in the regime above 1018 eV are currently best observed via radio detection of Askaryan radiation. A promising approach for the reconstruction of detected neutrino events uses convolutional neural networks, the training data for which is provided by the Monte Carlo simulation tool NuRadioMC. The following work presents an analysis of the features found in the simulated data, with the purpose of understanding the sensitivity of the deep-learning approach towards flavour identification and inelasticity measurements. Two data sets at different neutrino energy ranges of 1018 eV < Eν < 1018.1 eV and 1018.9 eV < Eν < 1019 eV were studied with an emphasis on the signal strength and the pulse time differences between the electromagnetic and hadronic showers. It was found that multiple factors including the viewing angle, the distance between detector and interaction vertex, and the strength of the LPM effect play a role in the signal strength seen in the radio antennas. A theoretical model for the expected time delays between the electromagnetic and hadronic showers was developed, showing that while the delay has a dependency on thevertex position of the interaction and the propagation direction of the neutrino, other factors such as the variation of theice density also have a significant impact on the signal arrival times. It was shown that visible and distinguishable signals for the electromagnetic and hadronic showers are expected for a large fraction of charged-current events.