Xinnan Ma, Pinchun Wang, Jingyou Ma, Yeting Lu, Shengjie Pan, Xiaowei Hu
Interictal epileptiform discharges (IEDs) are established electroencephalographic biomarkers that support epilepsy diagnosis and characterization, but their visual identification is time-consuming and subject to inter-rater variability. This review critically examines automated IED analysis from rule-based systems to contemporary deep learning, with particular attention to task definition, validation design, and clinical workflow. We use the term "accuracy paradox" as an author-defined descriptive label-not as an established field-wide concept-for the mismatch between high performance on presegmented epochs or recording-level classification and the temporal and spatial precision required for event-level interpretation. Accuracy, area under the curve, F1 score, concordance, and false positives per hour quantify different aspects of performance and should not be treated as interchangeable. Clinical evaluation should therefore specify the target event, temporal and spatial matching rules, recording duration, annotated event count, validation level, and intended use, while reporting sensitivity, precision, F1 score, false-positive burden, and workflow outcomes as appropriate. We also discuss neurodevelopmental and cellular mechanisms that may influence network excitability, including neuroinflammation, complement-associated synaptic remodeling, and chloride homeostasis. Available evidence supports biological plausibility but does not establish that AI-derived IED morphology can reveal molecular states in individual patients. We propose a translational framework based on independent multicenter validation, context-specific operating thresholds, robust evaluation of explanations, and human-in-the-loop triage. Prospective studies integrating scalp or intracranial EEG with spatially matched surgical tissue and single-cell or spatial transcriptomic profiling could test macro-to-molecular hypotheses, but these applications remain investigational.