Magnetic resonance electrical properties tomography (MREPT) is a promising technique for non-invasive, early cancer detection based on the electrical properties of tissues. However, noise sensitivity and artifact problems have historically hindered the clinical application of MREPT. In order to address these limitations, researchers have investigated several approaches, including data-driven neural network methods, modified analytical models, or a combination of both. However, these methods face the challenge of balancing the trade-off between reducing artifacts and maintaining tissue contrast in MREPT. This study presents an analysis of our proposed Reconstruction Error Compensation Neural Network (REC-NN) and provides new insights into the input features utilized in the model. Special focus was put on the Laplacian of the source signal for MREPT, which is the most important term in most analytical models, to make clear its role in REC-NN with regard to the shape, contrast and accuracy of the reconstructed conductivity map.