Ventricular tachycardia (VT), often caused by post-infarction re-entrant circuits, remains a leading cause of sudden cardiac death. In cases where VT cannot be induced or sus- tained, clinicians rely on sinus rhythm mapping to identify abnormal electrograms (EGMs) as surrogates for critical isthmus localization. However, interpreting voltage and activation maps, along with large numbers of EGMs, is challenging due to signal complexity and spa- tial variability. In this study, we propose a novel approach leveraging the Joint Time-Vertex Scattering Transform (JTVST), a recent advancement in Graph Signal Processing (GSP), to extract meaningful spatiotemporal features from EGMs. This method models the catheter's geometry as a graph and applies wavelet scattering transforms across both time and space to capture complex signal dependencies. We evaluate our method on seven patient datasets acquired with the Carto mapping system, using expert annotations as ground truth. Fea- tures derived from JTVST are used to train LightGBM classifiers under a Leave-One-Out cross-validation scheme. We compare performance against traditional clinical features (e.g., voltage, fractionation), engineered features, and state-of-the-art time-series feature extractors such as ROCKET, combined with a LightGBM classifier. Our method significantly improves precision over clinical baselines (> +15%), while maintaining comparable recall and speci- ficity, thus outperforming existing approaches in terms of F1-score (from 0.59 to 0.66) and ROC-AUC (from 0.87 to 0.91). On an external test set (6 cases), our approach also improves spatial precision, reducing Chamfer distance from 10.32 mm to 6.28 mm (-39%), indicating better alignment between detected abnormalities and clinical ablation targets. These results suggest that joint time-vertex analysis offers a promising and interpretable framework for VT substrate mapping and ablation planning.