A segregating PTK2B variant in a primary biliary cholangitis (PBC) family induces PBC-like autoimmune features in knock-in mice.

Focal cortical dysplasia (FCD) is one of the most common causes of medically refractory epilepsy (). Characterized by localized malformations in cortical development, FCD often necessitates surgical resection to achieve seizure freedom in refractory cases. The success of epilepsy surgery is critically dependent on precise localization of the seizure onset zone (SOZ), which involves determining both the hemisphere (lateralization) and the anatomical lobe (localization) from which seizures originate (). Traditionally, SOZ localization is achieved through a multi-modal evaluation process involving non-invasive modalities such as scalp EEG, MRI, PET, and SPECT. Intracranial electroencephalography (iEEG) is then employed as a confirmatory tool to refine and validate the presumed seizure focus prior to surgical intervention. While iEEG provides high-resolution spatiotemporal information, it is limited by its partial brain coverage due to the invasive nature of electrode implantation. Furthermore, visual interpretation of EEG recordings by experienced epileptologists is inherently subjective and time-intensive, often requiring years of clinical expertise. In an era of increasing demand for standardized and reproducible clinical tools, the integration of machine learning (ML) into the presurgical evaluation workflow is of growing interest (). While previous studies have demonstrated the feasibility of using interictal EEG for SOZ localization, such approaches rely on indirect biomarkers such as interictal epileptiform discharges (IEDs), connectivity alterations, or spectral imbalances (; ; ). These interictal markers may provide useful clues but generally reflect the broader irritative zone rather than the SOZ, and their spatial specificity can be limited—especially in FCD, where interictal abnormalities are often diffuse or non-specific. In contrast, ictal EEG reflects the actual electrophysiological dynamics of seizure initiation and provides more direct and specific information about the SOZ. Thus, ictal EEG generally offers greater value for precise localization compared to interictal recordings, particularly in complex pathologies such as FCD. Recent advances in ML have enabled data-driven approaches to SOZ classification using both interictal and ictal scalp EEG recordings. Numerous studies have employed supervised learning algorithms, such as support vector machines, random forests, and deep neural networks, to infer SOZ lateralization or localization based on either handcrafted features or learned signal representations (; ). Interictal-based models often leverage features such as IED rate, high-frequency oscillations, or functional connectivity, whereas ictal-based models focus on time-frequency evolution, spatial propagation, or seizure activity tracking from scalp EEG. Recent deep learning studies have further demonstrated the feasibility of automated seizure activity tracking and onset-zone localization using non-invasive scalp EEG recordings, highlighting the potential value of modeling spatiotemporal ictal propagation (). Despite these promising advances, existing approaches continue to face challenges in interpretability, patient-level generalizability, cohort specificity, and clinical integration. Moreover, many prior models rely on static features or time-averaged representations, potentially overlooking the dynamic nature of ictal activity (; ). In contrast, our framework explicitly captures the temporal evolution of frequency-domain patterns by applying time-windowed segmentation and extracting features across multiple frequency bands. This time-resolved approach enables a more nuanced characterization of ictal onset dynamics, enhancing both localization accuracy and clinical interpretability.