Document Type : Original/Review Paper

Authors

• Mohammad Nahvi ¹
• GholamReza Ardeshir ²
• Mehdi Ezoji ²
• Abbas Tafakhori ³
• Sajjad Shafiei ⁴

¹ Department of Electrical and Computer Engineering, Babol Noshirvani University of Technology, Babol, Iran

² Department of Electronics, Faculty of Electrical and Computer Engineering, Babol Noshirvani University of Technology, Babol, Iran

³ Department of Neurology, Imam Khomeini Hospital Complex, Tehran University of Medical Sciences, Tehran, Iran

⁴ Neurosurgery Department, Imam Hospital, Mazandaran University of Medical Sciences, Sari, Iran

Abstract

Accurate localization of the seizure onset zone (SOZ) using intracranial EEG (iEEG) is important for treatment planning in drug-resistant epilepsy. However, clinical SOZ markings and treatment-associated contact annotations may be discordant. This retrospective study evaluated an ambiguity-aware temporal-connectivity framework in 18 patients with temporal lobe epilepsy treated by resection (n=13) or ablation (n=5), all with Engel class I outcomes. Analysis was restricted to a clinically guided candidate region within 50 mm of at least one seizure-specific clinically marked SOZ contact. Time-resolved frequency-domain Granger-causality outflow maps were encoded using a bidirectional long short-term memory network with an ambiguity-aware three-class output and an auxiliary treatment-association objective. Primary evaluation used leave-one-seizure-out (LOSO) cross-validation across 73 seizures and three predefined scoring scenarios. Across three LOSO executions, Scenario A achieved 92.9±0.4% accuracy and 57.2±2.1% sensitivity, while patient-level aggregation in Scenario B achieved 89.2±1.6% accuracy and 89.8±1.8% sensitivity. Scenario C, using a treatment-aligned asymmetric scoring policy, achieved 84.6±1.7% accuracy, 95.9±0.7% sensitivity, 84.0±2.4% specificity, 63.3±3.4% precision, and 76.2±2.3% F1-score. Complementary leave-one-patient-out analysis showed lower patient-exclusive transfer performance, with Scenario C sensitivity and F1-score of 70.5±5.4% and 40.0±1.8%, respectively. Auxiliary supervision increased sensitivity while reducing specificity, whereas the capacity-matched unidirectional LSTM showed similar overall performance. Sensitivity analyses showed comparable performance for MVAR orders 5 and 10, with lower performance at order 15 and after reducing the retained frequency range from 1–100 to 1–50 Hz. Overall, SOZ-localization performance depends on connectivity representation, validation design, ambiguity policy, and supervision strategy.

Keywords

• intracranial EEG
• seizure onset zone
• Granger causality
• BiLSTM
• ambiguity-aware learning

Main Subjects

• H.3.15.2. Computational neuroscience