Automated detection of epileptic seizures from scalp electroencephalography (EEG) remains challenging because of strong inter-patient variability. This paper proposes an interpretable, multi-domain feature engineering framework for seizure detection on the CHB-MIT Scalp EEG Database and evaluates it under strictly patient-level protocols. In addition to conventional time-domain, frequency-domain, wavelet and nonlinear descriptors, each EEG segment is passed through a Leaky Integrate-and-Fire neuron, a biologically inspired spike-domain transformation from which eight spike-train statistics are extracted. Gradient-boosted classifiers are evaluated on six entirely unseen test patients, in a 20-patient Leave-One-Subject-Out (LOSO) analysis, and under both a class-balanced and the natural, unbalanced class distribution. XGBoost reaches a ROC-AUC of 0.796 on unseen test patients (0.803 after Optuna tuning) and a mean LOSO ROC-AUC of 0.781, which remains same under the natural class distribution. Paired ablation studies show that every feature domain carries seizure-related information on its own, but that the time-domain family, dominated by peak-to-peak amplitude, is the main source of transferable information: its removal reduces LOSO ROC-AUC, whereas the SNN features, despite a standalone ROC-AUC, add no measurable contribution once conventional features are present. Restricting the positive class to the original seizure annotations increases LOSO ROC-AUC by 0.055, indicating that the extended seizure-associated labelling yields a conservative estimate. Feature-level SHAP analysis identifies amplitude and alpha/beta power markers as the most stable predictors across patients. Together, these results provide a leakage-controlled and interpretable account of what multi-domain EEG features can and cannot contribute to cross-patient seizure detection.