Introduction. Ventricular fibrillation (VF) is a life-threatening arrhythmia characterized by chaotic electrical activity and rapid, irregular ventricular contractions. Its manifestation can vary under different pharmacological interventions, highlighting the importance of characterizing drug-specific effects on VF dynamics. Understanding these effects is critical for advancing personalized antiarrhythmic therapy. Methods. We propose a deep learning framework that integrates wavelet-based representations with time-aware modeling to classify VF episodes induced by amiodarone, diltiazem, flecainide, or no drug. Using electrocardiogram recordings from anesthetized dogs, we extract features via Continuous Wavelet Transform (CWT) and Scattering Wavelet Transform (SWT). These features are processed by Long Short-Term Memory networks designed to capture temporal patterns within each VF episode. Experiments and Results. We evaluated classification performance across three temporal VF segments, specifically beginning, middle, and end, to visualize the learned representations using Uniform Manifold Approximation and Projection. Results show that class separability improves as the VF episode progresses. SWT-based models consistently outperform CWT, achieving up to 68% macro F1-score on an independent test set. Amiodarone and flecainide overlap, pointing to similar electrophysiological effects. Conclusion. Our findings demonstrate that interpretable, time-aware deep learning models can effectively capture drug-specific VF dynamics. This approach offers promising potential for improved arrhythmia classification and drug response assessment.