Ventricular arrhythmia (VA) is the phenomenon most observed leading to the sudden cardiac arrest, which is one of the leading causes of cardiovascular deaths. Although the post-infarction scar on the myocardium was recognised as the main substrate of the arrhythmic event, there is, yet, to be a standard consensus for the characterisation of the arrhythmogenic scars. In this work, we wish to study whether the AI classification model trained on a imaging dataset was able to learn the underlying electrical mechanism of VA through the interpretability method and can be used to pinpoint the arrhythmogenic region. We proposed the edge-weighted graph convolutional network for VA classification, which based its prediction based on the LV wall thickness (WT) extracted from the CT imaging. The model was trained and evaluated on a large retrospective dataset of 600 CT images, and achieved the evaluation score of \(83.2\%\) , \(80.4\%\) and \(84.2\%\) of accuracy, sensitivity and specificity, respectively. Finally, we performed the evaluation of the model interpretability (LIME and integrated gradients) using the data from 10 additional patients for which the electro-anatomical mapping and inducibility VA data were available. Quantitative and visual assessment were done to compare the interpretability output with the reentry channel, manually segmented from the recorded activation map. The quantitative assessment based on the AHA 17-segment model showed the best interpretability method (LIME with quickshift segmentation) could achieve up to \(84.5\%\) , \(90.0\%\) , \(78.9\%\) of balanced accuracy, sensitivity and specificity, respectively. The visual assessment showed a strong correlation between the high coefficient regions of the interpretability outputs and the reentry channel among 7 out of 10 cases.