IntroductionDrug-resistant epilepsy (DRE) presents a critical clinical challenge, and accurate epileptogenic zone (EZ) localization is essential for successful surgical intervention. Stereoelectroencephalography (SEEG) is widely applied in clinical practice, but conventional visual inspection and high-frequency oscillations (HFOs) suffer from low efficiency, poor reproducibility, and insufficient representation of complex neural dynamics, failing to improve postsurgical seizure-free rates substantially. To address the limitations, we developed an automated EZ localization framework leveraging ictal SEEG high-resolution time-frequency representation (TFR) and convolutional neural network (CNN).MethodsTwenty three adult individuals with DRE who achieved Engel Class I surgical outcomes were enrolled in this study. The TFRs of ictal SEEG (0.5–500 Hz) were generated using Short-Time Fourier Transform (STFT), Continuous Wavelet Transform (CWT), and Superlet Transform, then were classified by VGG16/19 and ResNet-18/34/50.ResultsUnder subject-independent cross-validation, the optimal HFO-based method achieved an accuracy of 66.10 ± 1.40%, an F1-score of 49.49 ± 2.48% and an AUC of 67.95% ± 2.35%. By comparison, Superlets-VGG19 model achieved significantly better performance: an accuracy of 80.46 ± 4.61%, an F1-score of 79.52 ± 6.88% and an AUC of 80.06 ± 5.23%. Notably, the framework exhibited superior performance in the radiofrequency thermocoagulation (RFTC) sub-cohort, with an accuracy of 82.31 ± 3.41%, an F1-score of 82.56 ± 4.62%, and an AUC of 82.85 ± 3.80%.DiscussionThe proposed framework outperforms conventional HFO-based baselines and provides a reliable and objective tool for clinical EZ localization, and holds substantial optimization potential via integration of complex architectures such as Transformer and Mamba-based models.