AI enabled decision support systems in epilepsy surgery a scoping review

AI enabled decision support systems in epilepsy surgery a scoping review | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Article AI enabled decision support systems in epilepsy surgery a scoping review Kai Yu, Shuang Zhou, Meijia Song, Zaifu Zhan, Yu Hou, Yiran Song, and 7 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-8612799/v1 This work is licensed under a CC BY 4.0 License Status: Under Review Version 1 posted 10 You are reading this latest preprint version Abstract Artificial intelligence is increasingly explored to support decision-making in epilepsy surgery, yet evidence for implementation across the epilepsy surgery pathway remains limited. We conducted a scoping review of 145 studies published between January 2018 and May 2025 to map AI-enabled decision support systems across surgical stages and clinical tasks, characterize datasets by modality, size, geographic provenance and accessibility, and synthesize modeling practices, external validation and workflow integration. The literature is heavily concentrated in the pre-operative stage, with no included intra-operative studies and relatively few post-operative applications. Most studies rely on small, single-center and non-public datasets and use supervised CNN-based models. External validation and workflow-integrated evaluation are uncommon, and only a minority of systems report semi-integrated clinical workflows. These findings highlight key gaps in generalizability, workflow readiness and equity, and inform priorities for multi-center data resources, rigorous cross-site evaluation and clinically meaningful endpoints to enable safe, scalable adoption. Biological sciences/Computational biology and bioinformatics Health sciences/Health care Physical sciences/Mathematics and computing Health sciences/Medical research Scientific community and society/Scientific community Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Figure 6 Figure 7 Introduction Epilepsy surgery is a high-risk, resource-intensive, multidisciplinary care pathway that requires coordinated, time-sensitive decisions across pre-operative, intra-operative, and post-operative stages. Epilepsy affects more than 50 million people worldwide 1 , and global epidemiological estimates suggest that approximately 10.1 million people living with epilepsy are potential surgical candidates, with around 1.4 million new surgically treatable cases each year 2 . Despite its promise for drug-resistant epilepsy, surgical success depends on integrated presurgical evaluation, precise operative execution, and careful post-operative monitoring to maximize seizure control while minimizing the risk of neurological complications 3 – 5 . Because workflows, expertise, and data infrastructures vary across centers, achieving consistent decision-making and scalable implementation remains a health-system challenge. Across the pre-operative, intra-operative, and post-operative phases of epilepsy surgery, clinicians rely on a wide range of diagnostic modalities, including magnetic resonance imaging (MRI), positron emission tomography (PET), computed tomography (CT), scalp electroencephalography (EEG), and intracranial EEG, to support multidisciplinary decisions such as lesion detection, seizure-onset (SOZ) and epileptogenic zone (EZ) localization, electrode navigation, and postoperative outcome assessment. However, these decisions rarely rely on any single modality. Instead, they depend on how multimodal data are synthesized and interpreted, a process that remains inherently complex and variable. Interpretation is often influenced by center-specific practices, and clinician experience, leading to substantial inter-institution variability even when similar diagnostic modalities are available. This variability complicates standardization and limits the scalability of consistent decision support across epilepsy centers. Beyond interpretive subjectivity, epilepsy surgery workflows are further challenged by the substantial specialist time and expertise required to interrogate complex diagnostic data. Depending on the clinical question, clinicians may need to review hours to weeks of scalp or intracranial EEG recordings, scrutinize high-resolution structural MRI for subtle cortical abnormalities, interpret regional metabolic patterns on PET, or evaluate post-operative structural changes on CT. These datasets are high-dimensional, noisy, acquired under heterogeneous protocols, and demand specialized expertise that varies across clinicians and centers. As a result, manual analysis is not only time-intensive but also vulnerable to inter-observer variability, limiting standardization and scalability in surgical decision-making. These constraints motivate workflow-integrated, scalable decision-support systems that can distill clinically meaningful features and enable more reproducible evaluation across all stages of epilepsy surgery. With the rapid development of artificial intelligence (AI), an increasing number of studies have explored AI-enabled decision support across the epilepsy surgery pathway. Most of this literature is based on deep learning methods. These approaches have shown promise in tasks such as structural MRI–based lesion detection 6 – 8 , intracranial EEG–based SOZ/EZ localization 9 – 11 , automated functional mapping 12 , 13 , surgical planning and navigation 14 – 16 , and surgical outcomes prediction 17 – 20 . A small number of studies have also begun to investigate foundation model approaches, for example large language models (LLMs) for clinical text interpretation 21 . However, the literature remains largely focused on technical feasibility in controlled settings. Many systems remain offline research prototypes, some are developed as standalone decision-support tools 22 , 23 , and a small number have been partially integrated into existing clinical workflows 24 . Health systems still lack the evidence needed to decide whether and how these tools can be deployed. Cross-center validation and evaluations of workflow integration, patient impact, and safety are still uncommon. Current research on AI-enabled decision support in epilepsy surgery remains fragmented across clinical tasks, data modalities, and study designs, making it difficult to judge where evidence is mature enough for implementation across the full care pathway. Existing review articles have largely reflected this fragmentation. Many focus on a single component of the surgical workflow or on a single modality, which limits the ability of clinicians and health systems to compare approaches, prioritize prospective evaluation, and plan workflow integration. For example, several reviews concentrate on EEG or intracranial EEG analysis 25 – 30 , providing insights into electrophysiology-based SOZ/EZ localization but offering limited discussion of imaging-based approaches or post-operative evaluation. Conversely, neuroimaging-focused reviews centered on MRI, PET, or CT often exclude electrophysiological studies entirely, resulting in an incomplete and modality-isolated picture of current evidence 31 – 33 . This separation across modalities and tasks prevents a pathway level assessment of AI systems, including how they are validated, evaluated for clinical impact, and moved toward workflow integration. Overall, a unified synthesis through a health-systems lens is needed to map the distribution of evidence, characterize methodological choices, and identify translational gaps that constrain real-world adoption. To address these gaps, we focus on five guiding questions: (RQ1) How are AI approaches distributed across the surgical pathway, including their targeted clinical tasks and the data modalities used? (RQ2) What are the characteristics of the datasets underpinning these studies, including their geographic origins, sample sizes, and accessibility? (RQ3) What modeling approaches and training strategies have been adopted, and what evidence is reported to support generalizability, including external validation? (RQ4) How is model performance assessed and reported, and how far have these systems advanced toward workflow integration and real-world deployment? (RQ5) How have research trends evolved over time, including global participation and institutional contributions, and what do they suggest about the maturity of the field? To answer these questions, this scoping review provides a comprehensive and structured examination of AI-enabled decision support, primarily deep learning-based systems, across the epilepsy surgery pathway. We first summarize the clinical context of existing studies, including their surgical stages, targeted clinical tasks, and data modalities. We then characterize the underlying datasets by assessing their geographic sources, sample sizes, and accessibility. Next, we synthesize modeling approaches and training strategies, with particular attention to evidence supporting generalizability, including external validation practices. We also examine how performance is assessed and reported, and how far systems have advanced toward workflow integration and real-world deployment. Finally, we review publication patterns and institutional contributions to contextualize the evolution of the field. Together, these analyses map the current landscape, clarify implementation-relevant evidence gaps, and highlight priorities for developing and evaluating AI systems that can be adopted in real-world care pathways and across centers. Box 1 | Terms and Concepts Large Language Models (LLMs) : Large-scale neural networks trained on massive text corpora to acquire general language understanding and reasoning capabilities. LLMs can interpret clinical narratives, summarize reports, assist decision-making, and serve as generalizable backbones for medical tasks. LLM Prompting : Providing tailored instructions, demonstrations, or constraints to an LLM to guide its output toward a target task without fine-tuning. Includes zero-shot, few-shot, and chain-of-thought. Retrieval-Augmented Generation (RAG) : A framework that combines external knowledge retrieval with generative modeling, enabling models to produce responses that are accurate, up-to-date, and grounded in verifiable sources. Foundation Models (FMs) : Large-scale pre-trained models that learn generalizable representations across modalities or tasks, providing adaptable backbones for downstream medical applications. Transfer Learning : Fine-tuning models pre-trained on large datasets (e.g., ImageNet MRI models) for epilepsy-specific tasks. Self-supervised Learning (SSL) : Training models using proxy tasks (e.g., contrastive learning, masked modeling) to learn robust representations from unlabeled neuroimaging or EEG data. Semi-supervised Learning : Training with both labeled and unlabeled data, often using pseudo-labeling or consistency regularization to mitigate label scarcity. Unsupervised Learning : Learning patterns without explicit labels—for example anomaly detection, clustering, or autoencoder feature representation. Reinforcement Learning (RL) : Learning decision policies from reward signals, occasionally applied to trajectory planning or electrode optimization. Fusion Strategies : Definitions of early, intermediate, late, and hybrid fusion, describing how multimodal data are integrated. External Validation : Testing model generalizability on independent datasets from different centers or regions, categorized as single-center or multi-center external validation. Results Overview of the Scope This section provides an overview of the scoping review and summarizes its principal findings (Fig. 1). We synthesize the literature on AI-enabled decision support in epilepsy surgery across five implementation-relevant dimensions: clinical context along the surgical pathway, dataset characteristics, modeling approaches and evidence supporting generalizability, clinical integration, and research trends. As illustrated in Fig. 2 and Fig. 3, we first map AI applications across the surgical pathway by characterizing targeted clinical tasks and associated data modalities (RQ1) . We then examine the datasets used in these studies (Fig. 3), including their geographic origins, sample sizes, and levels of accessibility (RQ2) . Next, we synthesize modeling approaches and training strategies (Fig. 4), with attention to evidence reported to support generalizability, including external validation (RQ3) . We also review how performance is assessed and reported, and how far systems have advanced toward workflow integration and real-world deployment ( Fig. 4; RQ4) . Finally, in Fig. 5, we contextualize the evolution of the field by summarizing publication trends, study types, and regional contributions (RQ5) . Collectively, these results provide a pathway-level view of where evidence is concentrated and where implementation-relevant gaps remain, helping to prioritize future prospective evaluation and workflow integration efforts for adoption across centers. Study Selection and Characteristics Fig. 7 summarizes the study selection process. Of the 2,534 records initially identified, 1,299 duplicates were removed, leaving 1,235 records for title and abstract screening. At this stage, 854 records were excluded for not meeting the predefined inclusion criteria, resulting in 381 reports eligible for full-text review. During full-text screening, 236 records were further excluded due to the following reasons: out-of-scope setting or context (N=5), ineligible study design (N=54), or irrelevant methods (N=177). Ultimately, 145 studies were included in this scoping review. Across the included studies, evidence was heavily concentrated in the pre-operative stage (91%, N=132), with relatively few post-operative studies (8.3%, N=12), and one study spanning both stages. No included studies specifically addressed the intra-operative stage, highlighting a clear gap in pathway coverage. Within the pre-operative stage, the most common tasks were SOZ/EZ localization (60.7%, N=88), and lesion detection/classification (18.6%, N=27), followed by presurgical functional mapping (6.2%, N=9) and surgical planning and navigation (2.1%, N=3). Post-operative studies primarily focused on post-operative assessment and outcome prediction (7.6%, N=11), with one study addressing electrode localization/navigation support (0.7%, N=1). Data modalities varied by task, with electrophysiology data (scalp EEG, intracranial EEG, and MEG) commonly used for SOZ/EZ localization, and structural MRI predominantly used for lesion detection and classification. Dataset provenance showed broad geographic coverage, with 29.7% (N=43) using data from North America, 26.9% (N=39) from Asia, 26.9% (N=39) from Europe, and 3.5% (N=5) from Oceania. Multi-regional datasets were used in 12.4% of studies (N=18), reflecting an emerging trend toward broader data diversity. Modeling approaches were dominated by CNN-based methods (46.2%, N=67), followed by U-Net/FCN-based segmentation models (14.5%, N=21), with smaller proportions of hybrid CNN–RNN models (8.3%, N=12) and GNN-based architectures (5.5%, N=8). Two studies (1.4%) used large language models (LLMs), specifically ChatGPT-4, reflecting early exploration of foundation model approaches in this domain. Additional LLM-based studies 34,35 identified during screening were excluded because they focused on surgical candidate selection or pre-surgical triage rather than tasks within the epilepsy surgery pathway, consistent with our predefined criteria. Implementation readiness remained limited. Most studies were conducted as offline research (89.6%, N=130), a smaller proportion functioned as decision-support tools (7.6%, N=11), and only four studies (2.8%) were integrated into clinical systems. Publication output consisted of journal articles (74.5%, N=108) and conference papers (25.5%, N=37), with activity increasing over time, including notable increases in 2020 (16.6%, N=24) and 2022 (21.4%, N=31). Clinical Context As illustrated in Fig. 3a-c, we summarize the clinical contexts of the included studies, including their targeted clinical tasks, surgical stages, and data modalities. Among these studies, 91% (N=132) focused on the pre-operative stage, with the main clinical tasks involving SOZ/EZ localization, lesion detection/classification, presurgical functional mapping, and surgical planning and navigation. Notably, 60.7% (N=88) of all included studies centered on SOZ/EZ localization, making it the predominant research focus in current deep learning applications for epilepsy surgery. For example, recent studies have leveraged various electrophysiology modalities, including scalp EEG 23,36–43 , intracranial EEG 9–11,44–69 , and MEG 70–77 , to identify seizure-onset and epileptogenic zones. Other works have utilized MRI based data, such as structural MRI 78–80 and resting-state fMRI 81–85 , and two studies have analyzed clinical text 21,86 for SOZ/EZ localization. Beyond single modalities, several works also incorporated multimodal inputs. For instance, Li et al. 87 and Yang et al. 88 combined intracranial EEG and scalp EEG for SOZ/EZ localization. In addition, multiple studies such as Jeong et al. 89 , Fard et al. 90 , Banerjee et al. 91 , and Zotova et al. 92 employed various imaging modalities to perform SOZ/EZ localization, including MRI (T1, T2, FLAIR, and DWI) and PET. A total of 27 studies (18.6%) focused on lesion detection/classification task, most of which relied on structural MRI with multiple sequences 6–8,24,93–108 , Intracranial EEG 109,110 , pathology 111 , or multimodal inputs 112–114 for lesion detection/classification. Nine studies (6.2%) focused on presurgical functional mapping, and commonly used data modalities included diffusion MRI tractography 13,115–117 , motion visual analysis 12,118,119 , intracranial EEG 120 , and audio speech 121 . The remaining three studies (2.1%) focused on surgical planning and navigation. Granados et al. 14 proposed an early-fusion multimodal approach integrating T1-weighted MRI, DWI, and CT. Nejedly et al. 16 introduced a late-fusion multimodal method based on intracranial EEG and T1-weighted MRI. In contrast, Liu et al. 15 relied solely on T1-weighted MRI to perform surgical planning and navigation. Unlike the single-task studies summarized above, five studies addressed multiple pre-operative tasks. Hossain et al. 122 and Zhang et al. 123 used scalp EEG and intracranial EEG, respectively, to perform both SOZ/EZ localization and presurgical functional mapping. Mo et al. 124 employed an early-fusion multimodal approach integrating MRI sequences with T1, FLAIR and PET to achieve lesion detection/classification and surgical planning and navigation. Similarly, Park et al. 125 applied an early-fusion multimodal strategy, based on T1-weighted MRI and T2-FLAIR, to perform SOZ/EZ localization and lesion detection/classification. In addition, Hu et al. 126 used scalp EEG for both SOZ/EZ localization and surgical planning and navigation. In contrast, 8.3% (N=12) of studies addressed the post-operative stage, primarily targeting post-operative assessment and outcome prediction 17–20,22,127–132 and electrode localization/navigation support 133 . A variety of data modalities were applied in this stage, including diffusion tractography MRI 20,127 , structural MRI 22,128,129,131,132 , CT 130,133 , and electrophysiology (Intracranial EEG) 17,19 . Notably, Tang et al. 18 employed a multimodal late-fusion approach that integrated multi-sequence MRI (T1, T2, and FLAIR), CT, and PET to achieve post-operative assessment and outcome prediction. Additionally, one study examined both stages 134 , covering presurgical functional mapping in the pre-operative phase and post-operative assessment and outcome prediction in the post-operative phase. Overall, current AI-enabled decision-support research in epilepsy surgery remains heavily concentrated in the pre-operative stage, with SOZ/EZ localization and structural MRI–based lesion detection/classification accounting for most studies. In contrast, presurgical functional mapping, surgical planning and navigation, and post-operative evaluation remain relatively underrepresented, and no study in our review specifically addressed intra-operative applications. Although multimodal fusion has begun to appear in recent work, its adoption is still limited, and comprehensive multi-task models remain uncommon. The absence of intra-operative studies also indicates limited evidence for real-time, workflow-integrated decision support during surgery. Data Characteristics Across the included studies, 10 different data modalities and types were used. Electrophysiology was the most common (52.8%, N=84), primarily consisting of intracranial EEG 9–11,44–69 , scalp EEG 23,36–43 , and MEG 70–77 . This was followed by structural MRI (25.2%, N=40), mainly including T1-weighted, T2-weighted, and FLAIR MRI; diffusion MRI (6.9%, N=11), including DWI 13,14,20,89,91,115,116,134 , DTI 127 , and DKI 108 ; PET 90,124 (4.4%, N=7); functional MRI (rs-fMRI) 81,84 (3.8%, N=6); and CT 133 (2.5%, N=4). One study use histopathology for FCD classification 111 . Seven studies incorporated video 12,118,119 , text 21,86 , or audio data 121 . Notably, many studies did not rely on a single modality for epilepsy surgery analysis. Instead, they employed combinations of multiple data modalities, such as the fusion of T2-weighted and FLAIR MRI 98,99 , intracranial EEG with rs-fMRI 112 , intracranial EEG with scalp EEG 88 , and the integration of multiple MRI sequences with PET 114,124,135 or CT 14,18 . Collectively, the breadth of modalities and heterogeneous acquisition protocols highlight the need for data harmonization and interoperable pipelines to support cross-center deployment. Based on the geographic distribution of the datasets used in the included studies (Fig. 3d), most data originated from North America (29.7%, N=43), Asia (26.9%, N=39), and Europe (26.9%, N=39), with Oceania contributing 3.5% (N=5). In addition, 12.4% (N=18) of the studies utilized multi-regional datasets spanning two or more continents. For example, Yang et al. 88 combined intracranial EEG and scalp EEG data from Asia and Europe for pre-operative SOZ/EZ localization. Zhang et al. 135 similarly used data from Asia and Europe, but employed T1-weighted MRI, FLAIR MRI, and PET with a multimodal early-fusion strategy to achieve lesion detection/classification. Gill et al. 7 also adopted an early-fusion multimodal approach for lesion detection/classification, using T1-weighted MRI and FLAIR MRI collected from four regions: Asia, Europe, North America, and South America. The studies by Spitzer et al. 104 and Ripart et al. 106 incorporated the broadest geographic coverage to support lesion detection/classification, with T1-weighted and FLAIR MRI sourced from five regions (Asia, Europe, Oceania, North America, and South America). In contrast to the imaging- and electrophysiology-based studies above, Luo et al. 86 utilized clinical text collected from Asia, Europe, and North America to perform SOZ/EZ localization. Overall, the predominance of single-region datasets suggests that geographic representativeness and cross-center generalizability remain important constraints for implementation at scale. Most studies relied on private datasets (65.5%, N=95), spanning diverse data types including audio, video, text, CT, PET, and multiple MRI sequences. In contrast, 22.1% (N=32) used fully public datasets, of which 30 involved electrophysiology data, including scalp EEG and intracranial EEG, and two studies used structural MRI (T1-weighted and FLAIR). Another 12.4% (N=18) employed mixed datasets combining private and public sources, mainly covering text, electrophysiology, PET, and T1-weighted MRI. The dataset sizes exhibited substantial heterogeneity, ranging from single-subject datasets to large cohorts exceeding 2,000 patients. The cumulative distribution (Fig. 3e) shows a strongly right-skewed, long-tailed pattern, with 80% of the studies using fewer than 100 patients. Only a small proportion of studies 40,78,87,95,111,135 used medium-sized datasets (100–500 patients), and very few utilized large-scale cohorts 81,96,104 (>1,000 patients). Collectively, the dataset landscape is characterized by strong reliance on small, private, single-region datasets, with limited public data availability, which can impede external validation, reproducibility, and workflow-integrated evaluation needed for adoption across centers. Modeling Approaches To support AI-enabled decision support across diverse modalities, the included studies adopted a range of deep learning architectures (Fig. 4a). CNN-based models were the most common (46.2%, N=67), spanning 1D CNNs for electrophysiology 17,41,51,110,115,136 , 2D CNNs for MRI 84,85 , and 3D CNN for volumetric MRI 81,96,114,124 . In addition, 14.5% (N=21) used CNN-derived segmentation architectures, primarily U-Net or fully convolutional networks (FCNs), applied to PET 113,135 , CT 130,133 , and structural MRI data 8,15,22,95,98,100,102 . A further 8.3% (N=12) and 4.8% (N=7) of the studies adopted RNN-based 137–139 and hybrid CNN–RNN 140,141 architectures, respectively, most often for electrophysiology or video analysis. The remaining studies used a variety of model architectures, such as autoencoder-based 142,143 , Transformer-based 144 , and attention-based 68,70 models for specific tasks. Notably, one study 119 leveraged an existing detection framework (Mask R-CNN) to perform automated presurgical functional mapping from video data, and two studies 21,86 applied large language models (LLMs; ChatGPT-4) to clinical text for SOZ/EZ localization. For example, Luo et al. 86 addressed the SOZ/EZ localization task by using ChatGPT-4, which interpreted textual seizure semiology descriptions from literature and clinical records to automatically predict the most likely brain lobe of seizure onset. The training strategies were dominated by supervised learning, which accounted for 74.5% (N=108) of all models (Fig. 4c). Transfer learning was used in 8.3% (N=12) of the studies, typically through initialization with pretrained CNN backbones 18,37,118 . A smaller proportion employed semi-supervised learning 145–147 (5.5%, N=8) or unsupervised learning techniques 143,148 (4.1%, N=6), such as autoencoder-based feature learning 143 . Only a limited number of studies utilized more advanced training paradigms, such as pretraining followed by finetuning 56,69,74,75,129 (3.4%, N=5), multi-training schemes 9,117 (1.4%, N=2), or reinforcement learning 149 (0.7%, N=1). A small number of studies required no additional training 119 (0.7%, N=1) or relied purely on LLM prompting 86 (0.7%, N=1). One study additionally incorporated retrieval-augmented generation (RAG) for text-based clinical reasoning 21 (0.7%, N=1). Overall, these patterns suggest that most systems depend on labeled data and task-specific training, which may limit scalability when high-quality annotations are difficult to obtain. Evidence supporting generalizability remained limited (Fig. 4b). Most studies did not perform external validation (77.2%, N=112). A smaller subset conducted single-center external validation (17.2%, N=25), typically using an independent cohort from a different institution within the same geographic region. Several studies 23,57,65,86,88,104,135,145 validated models on external data from different geographic regions, providing a stronger test of cross-regional generalizability. Only 5.5% (N=8) of the studies 7,38,74,106,124,129,132,141 employed multi-center external validation, leveraging datasets from two or more independent sites to assess model generalizability across institutions. Taken together, while the field explores diverse architectures and training strategies, current epilepsy surgery AI is still dominated by supervised CNN-based approaches, and the scarcity of multi-center external validation indicates substantial gaps in implementation-relevant evidence for deployment across centers. Evaluation and Clinical Integration As shown in Fig. 4d, most studies assessed model performance using fully automatic evaluation procedures (97.2%, N=141), in which quantitative performance metrics were computed directly from algorithmic outputs without additional human involvement. Common metrics included classification measures such as accuracy, sensitivity, specificity, AUC, F1-score 40,78,84,144 , as well as segmentation measures such as IoU and DSC 7,106,135 . Only 1.4% (N=2) of the studies relied exclusively on manual evaluation. For example, Chiang et al. 21 developed an ontology-guided, GPT-based system to assist SOZ/EZ Localization from clinical text, and epilepsy specialists manually judged each model prediction as correct or incorrect to derive accuracy. Similarly, Pastore et al. 107 manually verified their algorithm generated lesion predictions against histopathology or scalp EEG findings. Another 1.4% (N=2) of studies employed mixed evaluation, combining automated quantitative metrics with expert review to adjudicate ambiguous cases or assess clinical plausibility. For instance, in Mo et al. 124 , sensitivity, specificity, and accuracy were automatically computed, but overlap validation and all trajectory feasibility and safety assessments were manually determined by clinical experts. With respect to clinical integration (Fig. 4e), most systems remained offline research prototypes (89.6%, N=130), without integration into routine clinical software environments or workflows. A smaller subset (7.6%, N=11) functioned as standalone decision-support tools 18,22,23,38,39,73,74,84,106,111,124 , in which clinicians accessed model outputs through independent research interfaces operating outside routine clinical systems such as PACS, SEEG analysis software, or platforms. In contrast, only 2.8% (N=4) of the studies reported semi-integrated clinical workflows, in which AI outputs were incorporated directly into interfaces already used in clinical practice, such as PACS viewers 107 , online SEEG review platforms 69 , or GUI-based electrophysiology analysis tools 150 , and prospectively validated MRI-based lesion detection systems applied in daily clinical reading 24 . These systems enabled clinicians to visualize, review, and interact with model predictions within existing workflows, but none were fully automated or embedded into hospital information systems. Taken together, performance evaluation remains predominantly algorithm-centric and automated, while evidence for workflow-integrated use is limited. Most systems remain offline research tools, and only a small minority report partial, clinician-in-the-loop integration, highlighting ongoing gaps in implementation readiness for deployment across centers. Publication Trends This section characterizes publication trends in epilepsy surgery AI research (Fig. 5). Overall, publication activity has increased substantially since 2018, with a notable rise between 2020 and 2022. Most studies were published as journals articles (74.5%, N=108), while conferences papers accounted for 25.5% (N=37) (Fig. 5a). Regionally, Asia and North America contributed the largest shares of publications (50.3%, N=73 and 32.4%, N=47, respectively). Followed by Europe (13.1%, N=19), Oceania (2.8%, N=4), and Africa (1.4%, N=2). At the country level, China produced the most studies (31.0%, N=45), followed by the United States (29.0%, N=42). India and Japan each contributed 7.6% (N=11), while France, the United Kingdom, and Canada contributed 4.8% (N=7), 3.4% (N=5), and 3.4% (N=5), respectively. The remaining countries collectively contributed 13.1% (N=19) of publications. Overall, these patterns indicate growing research activity and broadening international participation, while also highlighting uneven geographic distribution of evidence generation. Discussion Our results highlight substantial gaps in pathway coverage and implementation readiness for AI-enabled decision support in epilepsy surgery. Evidence is heavily concentrated in the pre-operative stage (91%), with only 8.3% of studies addressing post-operative evaluation and none specifically targeting intra-operative applications. This imbalance reflects differences in data accessibility, operational constraints, and clinical risk. For example, pre-operative EEG and MRI are routinely acquired with established protocols, whereas intra-operative data collection is technically challenging, time-sensitive, and often constrained by surgical workflow. Beyond stage imbalance, effective seizure focus localization often requires integrating structural imaging, electrophysiology, and seizure semiology. However, most studies focus on either imaging (e.g., MRI) or electrophysiology (EEG) in isolation, and comparatively few integrate these modalities with the video-based behavioral and semiology information. Meanwhile, intra-operative AI remains substantially underdeveloped despite its clear clinical importance. Future research should prioritize evidence generation for intra-operative, workflow-integrated decision support 120 , and develop pathway-level approaches that connect pre-operative, intra-operative, and post-operative stages through end-to-end or multi-stage pipelines. Within pre-operative studies, SOZ/EZ localization and lesion detection/classification dominate current research activity. These tasks often have relatively well-defined reference standards (e.g., SEEG-confirmed seizure onset zones, postoperative seizure outcomes), making them easier to benchmark. In contrast, tasks such as presurgical functional mapping and surgical planning/navigation more closely reflect real clinical decision making but remain relatively underexplored. This imbalance in task coverage may constrain clinical impact by leaving critical components of the care pathway insufficiently supported. Although multimodal analyses can offer advantages over single-modality approaches, few studies conduct systematic comparisons of how different fusion strategies, including early 7,14,87,100,101 , late 16,18,112 , or intermediate 113,151 fusion, affect performance and interpretability. Advancing this field will require standardized and interoperable multimodal pipelines, along with interpretable cross-modal alignment methods that integrate electrophysiology, structural and functional imaging, and clinical information into unified decision-support systems. Generalizability and implementation readiness are constrained by the current data landscape. Across included studies, dataset size represents a major limitation. 80% of studies used small cohorts with fewer than 100 patients, while large-scale datasets (>1,000 patients) were extremely rare. This reflects the inherent difficulty of collecting large, high-quality epilepsy surgery datasets, particularly those involving multimodal imaging or invasive electrophysiology 23,92,126 . However, reliance on small datasets increases the risk of overfitting and limits the reliability of model performance estimates 12,44,76 . To address these challenges, future research should leverage advanced methods such as self-supervised learning, semi-supervised learning, and foundation models to enhance representation learning and improve generalization under small dataset conditions. Geographic and institutional coverage is similarly uneven. Most datasets originate from North America, Asia, and Europe, and are predominantly single-center and not publicly available. Such single-region, single-center datasets may embed systematic biases related to patient demographics, scanner types, acquisition protocols, and surgical practices, thereby reducing the applicability of trained models to other populations or institutions. Moving forward, increased cross-regional collaboration and harmonization techniques such as domain adaptation and federated learning will be important to mitigate these biases 89,141 . Data accessibility represents an additional challenge. With 65.5% of datasets being private and unavailable for reuse, reproducibility is often limited, and fair method comparison remains challenging. Establishing open, multi-center, multimodal epilepsy surgery benchmarks, alongside broader sharing of pretrained models, and code, will be critical for accelerating progress and improving community confidence 68,93,129 . Beyond scale and accessibility, label quality also presents inherent limitations 145 . Ground-truth labels such as SOZ/EZ localization, FCD lesion masks, or HFO annotations frequently contain noise due to subjective expert interpretation, heterogeneous clinical standards, or reliance on proxy markers such as resection zones or short-term postoperative outcomes. Improved multi-rater consensus labeling, uncertainty modeling, and the use of long-term outcomes will be needed to strengthen the reliability of training and evaluation. Finally, data modality usage remains highly skewed. Electrophysiology modalities account for 52.8% of all datasets, whereas other rich information sources, such as PET, rs-fMRI, diffusion imaging, and especially clinical text, are rarely exploited. With recent advances in large language models, leveraging clinical narratives, reports, and semiology descriptions may open new opportunities for improving prediction accuracy and capturing patient-specific clinical context. Expanding beyond electrophysiology to more diverse data modalities will be an important direction for future work 21,118 . In terms of modeling, the field remains dominated by supervised CNN-based approaches. Most studies directly adapt existing architectures from general computer vision or signal processing, rather than models specifically tailored to the characteristics of multimodal neuroimaging, electrophysiology, or surgical decision-making. Although these models often produce strong results on small, single-center datasets, their robustness in multi-center, heterogeneous clinical environments remains largely unknown 6,106 . Moreover, there remains a notable absence of systematic comparisons across model families, such as CNNs, transformers, GNNs, hybrid architectures, or generative models, making it difficult to determine which design choices consistently translate across settings. Notably, large language models (LLMs) have recently been introduced into this field, primarily for interpreting clinical text for SOZ/EZ localization 21,86 . However, their application remains limited to unimodal textual reasoning. Future work should explore the potential of multimodal LLMs as decision-support agents capable of integrating imaging, electrophysiology, and clinical narratives. LLMs may also support auxiliary tasks, including EEG/MRI annotation, study design, and automated report generation. A common limitation is the lack of external validation. With 77.2% of studies reporting no external validation, performance estimates are likely inflated due to single-center bias. Strengthening cross-center and cross-region generalization should therefore be a central methodological priority. Multi-center external validation, domain adaptation, and harmonization strategies will be essential to ensure that models perform reliably across diverse patient populations and clinical environments 76,111 . Evaluation and translation evidence remain key bottlenecks for deploying epilepsy surgery AI in health systems. Most studies rely solely on automated quantitative metrics such as accuracy, AUC, or Dice scores. While these measures are essential for algorithmic benchmarking, they provide limited insight into whether a model improves clinically meaningful outcomes in epilepsy surgery. Incorporating clinically relevant endpoints, such as changes in surgical planning, accuracy of resection margin prediction, seizure-freedom rates, reduction in evaluation time, or improvements in clinician confidence, would provide a more comprehensive assessment of real-world utility 22,84 . Workflow integration is also uncommon. 89.6% of included studies operate as offline research prototypes, with only a small number developed into decision-support tools or semi-integrated systems. This translation gap reflects a broader disconnect between algorithm development and real clinical workflows, driven by challenges such as limited access to surgical environments, high barriers to software integration, compatibility issues with existing medical devices and IT infrastructure, unclear regulatory requirements, and the lack of collaboration between researchers, clinical engineers, and industry partners 80,84,120 . Moving forward, evaluations should measure real clinical impact rather than just accuracy, including effects on surgical planning, workflow efficiency, and clinician confidence. Furthermore, AI systems must be usable in real workflows, support clinician-in-the-loop interaction, and integrate smoothly with existing systems 73 . Publication volume in epilepsy surgery AI, predominantly deep learning-based, has increased since 2018 with notable accelerations in 2020 and 2022. This growth likely reflects the maturation of deep learning frameworks, increased availability of open-source tools and datasets, and broader adoption of computational methods within the neuroscience and medical imaging communities. Despite this momentum, the field still lacks standardized benchmarks or shared evaluation pipelines, which limits reproducibility and makes it difficult to compare methodological advances across studies. The geographic distribution of contributing research groups is similarly imbalanced. Asia and North America account for most publications, while contributions from Europe are moderate and those from other regions remain limited. This imbalance raises concerns regarding model fairness and global applicability, as datasets predominantly represent specific populations, acquisition protocols, and surgical practices 86 . Addressing these disparities will require coordinated multi-center consortia, greater international data sharing, and wider adoption of federated learning to support collaborative model development without requiring data exchange. These efforts are essential for building globally representative datasets and developing more generalizable and equitable AI-enabled decision-support systems for epilepsy surgery. Taken together, our findings indicate that epilepsy surgery AI is expanding but remains fragmented across the care pathway. Evidence is constrained by small, geographically skewed, and often non-public datasets, alongside limited external validation and scarce multi-center evaluation. Most models perform well in controlled research settings but remain untested in the heterogeneous, multi-center environments where real clinical decisions are made. Evaluation practices emphasize technical accuracy rather than clinical impact, and workflow-integrated deployment is uncommon. Addressing these gaps will require a transition toward full-pathway AI systems, large-scale multimodal datasets, advanced and interpretable modeling strategies, rigorous generalization testing, and stronger collaboration across technical, clinical, and industry domains. Fig. 6 summarizes these limitations and outlines future directions for developing reliable, equitable, and clinically actionable decision-support systems for epilepsy surgery. This survey also has several limitations. First, we restricted our search to English language publications, which may have excluded relevant studies reported in other languages. Second, our review focused on peer-reviewed articles and did not incorporate recent preprints, potentially missing emerging research trends. In summary, this scoping review provides a comprehensive and structured synthesis of AI-enabled decision support in epilepsy surgery. We examined how models are distributed across clinical stages and tasks, characterized the underlying datasets, analyzed modeling approaches and validation strategies, and assessed evaluation practices and workflow integration. We further identified key methodological, data-related, and translational limitations and outlined future directions needed to advance this field. Our findings highlight implementation-relevant gaps, including uneven pathway coverage, small and geographically concentrated datasets, limited external validation, and scarce workflow-integrated evaluation. We hope this work serves as a foundational reference for clinicians, researchers, and industry partners seeking to develop reliable, generalizable, and clinically impactful decision-support systems for epilepsy surgery. Methods Data Source and Article Selection This scoping review was reported in accordance with the PRISMA-ScR guidelines, and the study selection process is summarized in the PRISMA flow diagram (Fig. 7). The protocol for this scoping review was not prospectively registered. A systematic search was conducted across seven electronic databases, including Scopus, Web of Science, Embase, PubMed, IEEE Xplore, ACM Digital Library, and CINAHL, to identify studies on AI-enabled decision support in epilepsy surgery published between January 2018 and May 2025. The search strategy combined terms related to deep learning, epilepsy, and surgery. Specifically, search terms related to deep learning were "deep learning", "machine learning", "neural network*", "artificial intelligence", "AI", "LLM*", "large language model*", "foundation model*", "vision language model*", "vision-language model"; search terms related to epilepsy were "epilep*", and "seizure*"; and search terms related to surgery involved "surgery", "surgical", "presurgical", "preoperative", "intraoperative", "postoperative". Detailed search queries and Boolean operators (AND/OR) for each database are provided in the Supplementary Table 1. Study selection followed a two-stage screening process. First, two reviewers independently screened titles and abstracts using predefined inclusion and exclusion criteria. Second, full texts of potentially eligible studies were independently evaluated, with disagreements resolved through discussion or, if necessary, adjudication by a third reviewer. Specifically, studies were included if they met the following criteria: (1) empirical investigations of deep learning-based methods; (2) applications directly related to epilepsy surgery, including pre-operative, intra-operative, or post-operative stages. Conversely, studies were excluded if they: (1) relied solely on traditional machine learning methods; (2) focused on epilepsy-related applications that were not connected to the surgical pathway; (3) solely for surgical candidate screening or pre-surgical triage without involving tasks directly related to the operative workflow, or (4) were non-English publications, review articles, editorials, or abstracts without accessible full text. Data Extraction and Analysis Data from the included studies were systematically extracted using a predefined template to capture key information (Supplementary Table 2). Extracted items included: (1) basic study information, including title, publication type, published year, and author region (based on the first author’s primary institutional affiliation: Asia, Europe, North America, Oceania, or Africa); (2) epilepsy surgery stage & clinical task, specifying the targeted surgical stage (pre-operative, intra-operative, or post-operative), and the corresponding clinical task; (3) data characteristics, including data source region, data modality, data size, and dataset accessibility (public, private, or mixed). In this survey, data size was recorded as the number of patients whose data were used in the study, including both individuals with epilepsy and healthy controls; (4) modeling details, including deep model architecture and training strategy (e.g., supervised learning, transfer learning, self-supervised learning, weak supervision, federated learning, or other specified strategies); (5) evaluation, including external validation, evaluation type (manual, automatic, or mixed), reported evaluation metrics, and the primary performance summary; and (6) clinical translation, in which we categorized the deployment level of each deep learning system as offline research (models evaluated retrospectively without clinical use), standalone decision support tools (outputs available for clinician review but not embedded), or semi-integrated clinical systems (outputs partially incorporated into clinical workflows). All data extraction was performed independently by two reviewers, with discrepancies resolved through discussion and consensus. The basic information, clinical context, dataset characteristics, methodological details, and translational relevance of the included deep learning studies were systematically documented and categorized using our taxonomy. We synthesized these data to highlight key patterns in AI-enabled decision support systems for epilepsy surgery. We first summarized the survey scope, including surgery stage, clinical tasks, data modalities, and model architectures. We then analyzed study metadata such as geographic distribution, data accessibility, and training strategies. Finally, we reviewed evaluation practices and clinical translation stages and outlined major limitations and future directions. Declarations Data Availability All data generated during this study are presented in the Supplementary Materials. Code Availability No code was generated during this study. Acknowledgments This work was supported by the National Institutes of Health’s National Center for Complementary and Integrative Health under grant number R01AT009457, National Institute on Aging under grant number R01AG078154, and National Cancer Institute under grant number R01CA287413. The content is solely the responsibility of the authors and does not represent the official views of the National Institutes of Health. We thank Janie Sha and Yitao Cheng (Wayzata High School) for assistance with literature searching and reference organization. Author Contributions K.Y. and R.Z. conceptualized the study and led the work. K.Y. and S.Z. searched papers. K.Y., S.Z., M.S., Z.Z., Y.H, Y.S., M.Z., and B.Y. conducted paper screening and data extraction. K.Y. performed data synthesis and contributed to the writing. K.Y., F.L., S. P., Z. S., M.L., and R.Z. revised the manuscript. R.Z. supervised the study. All authors read and approved the final version. Competing Interests The authors declare no competing interests. References World Health Organization. Epilepsy: Key facts. World Health Organization https://www.who.int/news-room/fact-sheets/detail/epilepsy (2024). Vaughan, K. A. et al. An estimation of global volume of surgically treatable epilepsy based on a systematic review and meta-analysis of epilepsy. J. Neurosurg. 130 , 1127–1141 (2019). Thijs, R. D., Surges, R., O’Brien, T. J. & Sander, J. W. Epilepsy in adults. Lancet 393 , 689–701 (2019). Ryvlin, P., Cross, J. H. & Rheims, S. Epilepsy surgery in children and adults. Lancet Neurol. 13 , 1114–1126 (2014). Rosenow, F. & Lüders, H. Presurgical evaluation of epilepsy. Brain : a journal of neurology 124 Pt 9 , 1683–1700 (2001). Ito, Y. et al. Deep learning-based diagnosis of temporal lobe epilepsy associated with hippocampal sclerosis: An MRI study. Epilepsy Res. 178 , 106815 (2021). Gill, R. S. et al. Multicenter validation of a deep learning detection algorithm for focal cortical dysplasia. Neurology 97 , e1571–e1582 (2021). Niyas, S. et al. Segmentation of focal cortical dysplasia lesions from magnetic resonance images using 3D convolutional neural networks. Biomed. Signal Process. Control 70 , 102951 (2021). Daoud, H. & Bayoumi, M. Deep learning approach for epileptic focus localization. IEEE Trans. Biomed. Circuits Syst. 14 , 209–220 (2020). Xiao, L. et al. Automatic localization of seizure onset zone from high-frequency SEEG signals: A preliminary study. IEEE J. Transl. Eng. Health Med. 9 , 1–10 (2021). Saminu, S. et al. Application of deep learning and WT-SST in localization of epileptogenic zone using epileptic EEG signals. Appl. Sci. (Basel) 12 , 4879 (2022). Ahmedt-Aristizabal, D. et al. Deep facial analysis: A new phase I epilepsy evaluation using computer vision. Epilepsy Behav. 82 , 17–24 (2018). Lee, M.-H. et al. Novel deep learning network analysis of electrical stimulation mapping-driven diffusion MRI tractography to improve preoperative evaluation of pediatric epilepsy. IEEE Trans. Biomed. Eng. 67 , 3151–3162 (2020). Granados, A. et al. Towards uncertainty quantification for electrode bending prediction in stereotactic neurosurgery. in 2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI) 674–677 (IEEE, 2020). Liu, H., Holloway, K. L., Englot, D. J. & Dawant, B. M. A multi-rater comparative study of automatic target localization methods for epilepsy deep Brain Stimulation procedures. arXiv [eess.IV] (2022) doi:10.1117/12.2612927. Nejedly, P. et al. Leveraging interictal multimodal features and graph neural networks for automated planning of epilepsy surgery. Brain Commun. 7 , fcaf140 (2025). Constantino, A. C. et al. Expert-level intracranial electroencephalogram ictal pattern detection by a deep learning neural network. Front. Neurol. 12 , 603868 (2021). Tang, Y. et al. Machine learning-derived multimodal neuroimaging of presurgical target area to predict individual’s seizure outcomes after epilepsy surgery. Front. Cell Dev. Biol. 9 , 669795 (2021). Sun, L. et al. High-performance prediction of epilepsy surgical outcomes based on the genetic neural networks and hybrid iEEG marker. Sci. Rep. 14 , 6198 (2024). Lee, M.-H. et al. Deep learning-based tract classification of preoperative DWI tractography advances the prediction of short-term postoperative language improvement in children with drug-resistant epilepsy. IEEE Trans. Biomed. Eng. 72 , 565–576 (2025). Chiang, K.-L. et al. Customized GPT model largely increases surgery decision accuracy for pharmaco-resistant epilepsy. J. Clin. Neurosci. 130 , 110918 (2024). Arnold, T. C. et al. Deep learning-based automated segmentation of resection cavities on postsurgical epilepsy MRI. NeuroImage Clin. 36 , 103154 (2022). Sun, R., Sohrabpour, A., Worrell, G. A. & He, B. Deep neural networks constrained by neural mass models improve electrophysiological source imaging of spatiotemporal brain dynamics. Proc. Natl. Acad. Sci. U. S. A. 119 , e2201128119 (2022). Chanra, V. et al. Development and prospective clinical validation of a convolutional neural network for automated detection and segmentation of focal cortical dysplasias. Epilepsy Res. 202 , 107357 (2024). Zhang, X., Zhang, X., Huang, Q. & Chen, F. A review of epilepsy detection and prediction methods based on EEG signal processing and deep learning. Front. Neurosci. 18 , (2024). Si, Y. Machine learning applications for electroencephalograph signals in epilepsy: a quick review. Acta Epileptologica 2 , (2020). Nafea, M. S. & Ismail, Z. H. Supervised Machine learning and Deep learning techniques for epileptic seizure recognition using EEG signals-A systematic literature review. Bioengineering (Basel) 9 , 781 (2022). Rasheed, K. et al. Machine Learning for Predicting Epileptic Seizures Using EEG Signals: A Review. IEEE Rev. Biomed. Eng. 14 , 139–155 (2021). Qamar, W. U. R., Lee, M.-H. & Abibullaev, B. Deep learning in intracranial EEG for seizure detection: advances, challenges, and clinical applications. Frontiers in Neuroscience 19 , 1677898 (2025). Daida, A. et al. AI ‐based localization of the epileptogenic zone using intracranial EEG. Epilepsia Open (2025) doi:10.1002/epi4.70130. Berger, M., Licandro, R., Nenning, K.-H., Langs, G. & Bonelli, S. B. Artificial intelligence applied to epilepsy imaging: Current status and future perspectives. Revue Neurologique 181 , 420–424 (2025). Sone, D. & Beheshti, I. Clinical application of machine learning models for brain imaging in epilepsy: A review. Front. Neurosci. 15 , 684825 (2021). Yuan, J. et al. Machine learning applications on neuroimaging for diagnosis and prognosis of epilepsy: A review. Journal of Neuroscience Methods 368 , 109441 (2022). Tan, S. et al. Epilepsy surgery candidate identification with artificial intelligence: An implementation study. J. Clin. Neurosci. 135 , 111144 (2025). Chowdhury, S. et al. Using large language models to identify candidates for pediatric epilepsy surgery (P3-8.001). Neurology 102 , (2024). Hao, Y., Khoo, H. M., von Ellenrieder, N., Zazubovits, N. & Gotman, J. DeepIED: An epileptic discharge detector for EEG-fMRI based on deep learning. NeuroImage Clin. 17 , 962–975 (2018). Raghu, S., Sriraam, N., Temel, Y., Rao, S. V. & Kubben, P. L. EEG based multi-class seizure type classification using convolutional neural network and transfer learning. Neural Netw. 124 , 202–212 (2020). Abou Jaoude, M. et al. Noninvasive detection of hippocampal epileptiform activity on scalp electroencephalogram. JAMA Neurol. 79 , 614–622 (2022). Yu, Z. et al. Electrophysiological brain imaging based on simulation-driven deep learning in the context of epilepsy. Neuroimage 285 , 120490 (2024). Vajiram, J., Sivakumar, Jena, R. & Maurya, U. Epilepsy detection by different modalities with the use of AI-assisted models. Artificial Intelligence and Applications 2 , 233–246 (2023). Kuang, Z. et al. Seizure onset zone detection based on convolutional neural networks and EEG signals. Brain Sci. 14 , 1090 (2024). Dokare, I. & Gupta, S. Brain-region specific epileptic seizure detection through EEG dynamics: integrating spectral features, SMOTE and long short-term memory networks. Cogn. Neurodyn. 19 , 67 (2025). Liu, Y., Wang, Y. & Wang, T. Non-invasive localization of epileptogenic zone in drug-resistant epilepsy based on time-frequency analysis and VGG convolutional neural network. Bioengineering (Basel) 12 , 443 (2025). Medvedev, A., Agoureeva, G. I. & Murro, A. A Long Short-Term Memory neural network for the detection of epileptiform spikes and high frequency oscillations. Sci. Rep. 9 , (2019). Ma, K. et al. Automatic detection of high frequency oscillations (80-500Hz) based on convolutional neural network in human intracerebral electroencephalogram. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2019 , 5133–5136 (2019). Sui, L., Zhao, X., Zhao, Q., Tanaka, T. & Cao, J. Localization of epileptic foci by using convolutional neural network based on iEEG. in IFIP Advances in Information and Communication Technology 331–339 (Springer International Publishing, Cham, 2019). Bajaj, V., Taran, S., Tanyildizi, E. & Sengur, A. Robust approach based on convolutional neural networks for identification of focal EEG signals. IEEE Sens. Lett. 3 , 1–4 (2019). Daoud, H. & Bayoumi, M. Semi-supervised learning for epileptic focus localization using deep convolutional autoencoder. in 2019 IEEE Biomedical Circuits and Systems Conference (BioCAS) 1–4 (IEEE, 2019). Madhavan, S., Tripathy, R. K. & Pachori, R. B. Time-frequency domain deep convolutional neural network for the classification of focal and non-focal EEG signals. IEEE Sens. J. 20 , 3078–3086 (2020). Daoud, H. & Bayoumi, M. A convolutional gated recurrent neural network for seizure onset localization. in 2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 2572–2576 (IEEE, 2020). Zhao, X. et al. Classification of epileptic IEEG signals by CNN and data augmentation. in ICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 926–930 (IEEE, 2020). Fraiwan, L. & Alkhodari, M. Classification of focal and non-focal epileptic patients using single channel EEG and long short-term memory learning system. IEEE Access 8 , 77255–77262 (2020). Wang, Z. & Li, C. Classifying cross-frequency coupling pattern in epileptogenic tissues by convolutional neural network. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2020 , 3440–3443 (2020). Chung, Y. G. et al. Deep convolutional neural network based interictal-preictal electroencephalography prediction: Application to focal cortical dysplasia type-II. Front. Neurol. 11 , 594679 (2020). Zhao, X., Sui, L., Tanaka, T. & Cao, J. Epileptic Focus Localization Based on iEEG Plot Images by Using Convolutional Neural Network. Qin, H., Wu, M., Wan, X. & Du, Y. Stacked sparse autoencoder based automatic detection of ripples and fast ripples in epilepsy. in 2020 39th Chinese Control Conference (CCC) 2833–2837 (IEEE, 2020). Wang, Y. et al. Computer-aided intracranial EEG signal identification method based on a multi-branch deep learning fusion model and clinical validation. Brain Sci. 11 , 615 (2021). Xia, M., Sui, L., Zhao, X., Tanaka, T. & Cao, J. Convolution neural network recognition of epileptic foci based on composite signal processing of electroencephalograph data. Procedia Comput. Sci. 192 , 688–696 (2021). Yamamoto, S. et al. Data-driven electrophysiological feature based on deep learning to detect epileptic seizures. J. Neural Eng. 18 , 056040 (2021). Takayanagi, Y., Takayama, Y., Iijima, K., Iwasaki, M. & Ono, Y. Efficient detection of high-frequency biomarker signals of epilepsy by a transfer-learning-based convolutional neural network. Adv. Biomed. Eng. 10 , 158–165 (2021). Modak, S., Singha Roy, S., Bose, R. & Chatterjee, S. Focal epileptic area recognition employing cross EEG rhythm spectrum images and convolutional neural network. IEEE Sens. J. 21 , 23335–23343 (2021). Daoud, H. & Bayoumi, M. Generative adversarial network based semi-supervised learning for epileptic focus localization. in 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 1281–1285 (IEEE, 2021). Sui, L., Zhao, X., Zhao, Q., Tanaka, T. & Cao, J. Hybrid convolutional neural network for localization of epileptic focus based on iEEG. Neural Plast. 2021 , 6644365 (2021). Ren, G. et al. Identification of epileptogenic and non-epileptogenic high-frequency oscillations using a multi-feature convolutional neural network model. Front. Neurol. 12 , 640526 (2021). Guo, J., Wang, Y., Yang, Y. & Kang, G. IEEG-TCN: A concise and robust temporal convolutional network for intracranial electroencephalogram signal identification. in 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM) 668–673 (IEEE, 2021). Zhao, X., Solé-Casals, J., Zhao, Q., Cao, J. & Tanaka, T. Multi-feature fusion for epileptic focus localization based on tensor representation. in 2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC) 1323–1327 (IEEE, 2021). Wang, Y. et al. A cross-patient SEEG epileptic signal detection method based on adaptive feature fusion of brain network features and single-channel features. in Proceedings of the 2022 9th International Conference on Biomedical and Bioinformatics Engineering 165–172 (ACM, New York, NY, USA, 2022). Wang, Y. et al. Automatic localization of seizure onset zone based on multi-epileptogenic biomarkers analysis of single-contact from interictal SEEG. Bioengineering (Basel) 9 , (2022). Chen, J. et al. BrainNet: Epileptic wave detection from SEEG with hierarchical graph diffusion learning. in Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining 2741–2751 (ACM, New York, NY, USA, 2022). Zhao, X. et al. A multi-head self-attention deep learning approach for detection and recommendation of neuromagnetic high frequency oscillations in epilepsy. Front. Neuroinform. 16 , 771965 (2022). Zhang, M., Liu, J., Liu, C., Wu, T. & Peng, X. An efficient CADNet for classification of high-frequency oscillations in magnetoencephalography. in 2022 4th International Conference on Robotics and Computer Vision (ICRCV) (IEEE, 2022). doi:10.1109/icrcv55858.2022.9953255. Guo, J. et al. Transformer-based high-frequency oscillation signal detection on magnetoencephalography from epileptic patients. Front. Mol. Biosci. 9 , 822810 (2022). Zheng, L. et al. An artificial intelligence-based pipeline for automated detection and localisation of epileptic sources from magnetoencephalography. J. Neural Eng. 20 , (2023). Sun, R., Zhang, W., Bagić, A. & He, B. Deep learning based source imaging provides strong sublobar localization of epileptogenic zone from MEG interictal spikes. Neuroimage 281 , 120366 (2023). Guo, J. et al. A stacked sparse autoencoder-based detector for automatic identification of neuromagnetic high frequency oscillations in epilepsy. IEEE Trans. Med. Imaging 37 , 2474–2482 (2018). Liu, J. et al. A novel MEGNet for classification of high-frequency oscillations in magnetoencephalography of epileptic patients. Complexity 2020 , 1–9 (2020). Guo, J. et al. Automatic and accurate epilepsy ripple and fast ripple detection via virtual sample generation and attention neural networks. IEEE Trans. Neural Syst. Rehabil. Eng. 28 , 1710–1719 (2020). Park, K.-I. et al. Lateralizing value of artificial intelligence-based segmentation software in MRI-negative focal epilepsy. J. Epilepsy Res. 14 , 59–65 (2024). Kaestner, E. et al. Convolutional neural network algorithm to determine lateralization of seizure onset in patients with epilepsy: A proof-of-principle study: A proof-of-principle study. Neurology 101 , e324–e335 (2023). Flaus, A. et al. Deep-learning predicted PET can be subtracted from the true clinical fluorodeoxyglucose PET co-registered to MRI to identify the epileptogenic zone in focal epilepsy. Epilepsia Open 8 , 1440–1451 (2023). Luckett, P. H. et al. Deep learning resting state functional magnetic resonance imaging lateralization of temporal lobe epilepsy. Epilepsia 63 , 1542–1552 (2022). Nandakumar, N., Hsu, D., Ahmed, R. & Venkataraman, A. DeepEZ: A graph convolutional network for automated epileptogenic zone localization from resting-state fMRI connectivity. IEEE Trans. Biomed. Eng. 70 , 216–227 (2023). Nandakumar, N., Hsu, D., Ahmed, R. & Venkataraman, A. A deep learning framework to characterize noisy labels in epileptogenic zone localization using functional connectivity. Proc. IEEE Int. Symp. Biomed. Imaging 2024 , 1–5 (2024). Kamboj, P., Banerjee, A. & Gupta, S. K. S. Expert knowledge driven human-AI collaboration for medical imaging: A study on epileptic seizure onset zone identification. IEEE Trans. Artif. Intell. 5 , 5352–5368 (2024). Kamboj, P., Banerjee, A., Boerwinkle, V. L. & Gupta, S. K. S. The expert’s knowledge combined with AI outperforms AI alone in seizure onset zone localization using resting state fMRI. Front. Neurol. 14 , 1324461 (2023). Luo, Y. et al. Clinical value of ChatGPT for epilepsy presurgical decision-making: Systematic evaluation of seizure semiology interpretation. J. Med. Internet Res. 27 , e69173 (2025). Li, Z. et al. Deep learning of simultaneous intracranial and scalp EEG for prediction, detection, and lateralization of mesial temporal lobe seizures. Front. Neurol. 12 , 705119 (2021). Yang, Y., Li, F., Luo, J., Qin, X. & Huang, D. Epileptic focus localization using transfer learning on multi-modal EEG. Front. Comput. Neurosci. 17 , 1294770 (2023). Jeong, J.-W. et al. Multi-scale deep learning of clinically acquired multi-modal MRI improves the localization of seizure onset zone in children with drug-resistant epilepsy. IEEE J. Biomed. Health Inform. 26 , 5529–5539 (2022). Fard, A. S. et al. Image synthesis of interictal SPECT from MRI and PET using machine learning. Front. Neurol. 15 , 1383773 (2024). Banerjee, S. et al. Localization of epileptogenicity using incomplete MRI sequence data in children with seizure onset zone involving temporal lobe. in 2024 IEEE International Symposium on Biomedical Imaging (ISBI) vol. 2 1–5 (IEEE, 2024). Zotova, D., Jung, J. & Lartizien, C. GAN-based synthetic FDG PET images from T1 brain MRI can serve to improve performance of deep unsupervised anomaly detection models. in Simulation and Synthesis in Medical Imaging 142–152 (Springer International Publishing, Cham, 2021). Jiang, J. et al. Automated detection of hippocampal sclerosis using real-world clinical MRI images. Front. Neurosci. 17 , 1180679 (2023). Wang, H., Ahmed, S. N. & Mandal, M. Automated detection of focal cortical dysplasia using a deep convolutional neural network. Comput. Med. Imaging Graph. 79 , 101662 (2020). Zhang, D. et al. Deep learning and radiomics based automatic diagnosis of hippocampal sclerosis. Int. J. Neurosci. 133 , 947–958 (2023). Kaestner, E. et al. Adding the third dimension: 3D convolutional neural network diagnosis of temporal lobe epilepsy. Brain Commun. 6 , fcae346 (2024). Zhang, S. et al. Deep learning-based automated lesion segmentation on pediatric focal cortical dysplasia II preoperative MRI: a reliable approach. Insights Imaging 15 , 71 (2024). Bijay Dev, K. M. et al. Automatic detection and localization of Focal Cortical Dysplasia lesions in MRI using fully convolutional neural network. Biomed. Signal Process. Control 52 , 218–225 (2019). Feng, C., Zhao, H., Li, Y. & Wen, J. Automatic localization and segmentation of focal cortical dysplasia in FLAIR-negative patients using a convolutional neural network. J. Appl. Clin. Med. Phys. 21 , 215–226 (2020). Aminpour, A., Ebrahimi, M. & Widjaja, E. Lesion localization in paediatric epilepsy using patch-based convolutional neural network. in Lecture Notes in Computer Science 216–227 (Springer International Publishing, Cham, 2020). Alaverdyan, Z., Jung, J., Bouet, R. & Lartizien, C. Regularized siamese neural network for unsupervised outlier detection on brain multiparametric magnetic resonance imaging: Application to epilepsy lesion screening. Med. Image Anal. 60 , 101618 (2020). Aminpour, A., Ebrahimi, M. & Widjaja, E. Lesion Segmentation in Paediatric Epilepsy Utilizing Deep Learning Approaches. Adv. Artif. Intell. Mach. Learn 2 , 422–440 (2022). Karthika, A., Subramanian, R. & Karthik, S. Using a recurrent neural network with S2 characteristics, efficient identification of localised cortical dysplasia. J. Intell. Fuzzy Syst. 42 , 6293–6306 (2022). Spitzer, H. et al. Robust and generalisable segmentation of subtle epilepsy-causing lesions: A graph convolutional approach. arXiv [eess.IV] (2023) doi:10.1007/978-3-031-43993-3_41. Zhang, X. et al. Focal cortical dysplasia lesion segmentation using multiscale transformer. Insights Imaging 15 , 222 (2024). Ripart, M. et al. Detection of epileptogenic focal cortical dysplasia using graph neural networks: A MELD study. JAMA Neurol. 82 , 397–406 (2025). Pastore, L. V. et al. Integrating standard epilepsy protocol, ASL-perfusion, MP2RAGE/EDGE and the MELD-FCD classifier in the detection of subtle epileptogenic lesions: a 3 Tesla MRI pilot study. Neuroradiology 67 , 665–675 (2025). Huang, J., Xu, J., Kang, L. & Zhang, T. Identifying epilepsy based on deep learning using DKI images. Front. Hum. Neurosci. 14 , 590815 (2020). Zhao, B. et al. Integrated automatic detection, classification and imaging of high frequency oscillations with stereoelectroencephalography. Front. Neurosci. 14 , 546 (2020). Gupta, S., Meena, J. & Gupta, O. Neural network based epileptic EEG detection and classification. ADCAIJ abs/2111.03268 , (2020). Vorndran, J. et al. A deep learning-based histopathology classifier for Focal Cortical Dysplasia. Neural Comput. Appl. 35 , 12775–12792 (2023). Hosseini, M.-P., Tran, T. X., Pompili, D., Elisevich, K. & Soltanian-Zadeh, H. Multimodal data analysis of epileptic EEG and rs-fMRI via deep learning and edge computing. Artif. Intell. Med. 104 , 101813 (2020). Lu, Y. et al. A novel fusion mechanism for multimodal neuroimaging of focal cortical dysplasias. in Proceedings of the 2023 10th International Conference on Biomedical and Bioinformatics Engineering 28–37 (ACM, New York, NY, USA, 2023). Zhao, X. et al. Deep learning approaches for imaging-based automated segmentation of tuberous sclerosis complex. J. Clin. Med. 13 , 680 (2024). Xu, H., Dong, M., Nakai, Y., Asano, E. & Jeong, J.-W. Automatic detection of eloquent axonal pathways in diffusion tractography using intracanial electrical stimulation mapping and convolutional neural networks. in 2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018) 1034–1037 (IEEE, 2018). Xu, H. et al. Objective detection of eloquent axonal pathways to minimize postoperative deficits in pediatric epilepsy surgery using diffusion tractography and convolutional neural networks. IEEE Trans. Med. Imaging 38 , 1910–1922 (2019). Gruen, J., Bauer, T., Rüber, T. & Schultz, T. Deep learning based tractography with TractSeg in patients with hemispherotomy: Evaluation and refinement. NeuroImage Clin. 45 , 103738 (2025). Ahmedt-Aristizabal, D. et al. A hierarchical multimodal system for motion analysis in patients with epilepsy. Epilepsy Behav. 87 , 46–58 (2018). Ahmedt-Aristizabal, D. et al. Motion signatures for the analysis of seizure evolution in epilepsy. Annu. Int. Conf. IEEE Eng. Med. Biol. Soc. 2019 , 2099–2105 (2019). RaviPrakash, H. et al. Deep learning provides exceptional accuracy to ECoG-based functional language mapping for epilepsy surgery. Front. Neurosci. 14 , 409 (2020). Radha, K. & Narayana, S. DirichNet model for detection of TMS-induced speech errors in patients undergoing epilepsy surgery. in ICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 1–5 (IEEE, 2025). Hossain, M. S., Amin, S. U., Alsulaiman, M. & Muhammad, G. Applying deep learning for epilepsy seizure detection and brain mapping visualization. ACM Trans. Multimed. Comput. Commun. Appl. 15 , 1–17 (2019). Zhang, Y. et al. Characterizing physiological high-frequency oscillations using deep learning. J. Neural Eng. 19 , 066027 (2022). Mo, J. et al. Automated detection and surgical planning for focal cortical dysplasia with multicenter validation. Neurosurgery 91 , 799–807 (2022). Park, D. K. et al. Convolutional neural network-aided tuber segmentation in tuberous sclerosis complex patients correlates with electroencephalogram. Epilepsia 63 , 1530–1541 (2022). Hu, M. et al. E2SGAN: EEG-to-SEEG translation with generative adversarial networks. Front. Neurosci. 16 , 971829 (2022). Gleichgerrcht, E. et al. Deep learning applied to whole-brain connectome to determine seizure control after epilepsy surgery. Epilepsia 59 , 1643–1654 (2018). Pérez-García, F. et al. Simulation of brain resection for cavity segmentation using self-supervised and semi-supervised learning. in Lecture Notes in Computer Science 115–125 (Springer International Publishing, Cham, 2020). Pérez-García, F. et al. A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections. Int. J. Comput. Assist. Radiol. Surg. 16 , 1653–1661 (2021). Pantovic, A., Ren, X., Wemmert, C., Ollivier, I. & Essert, C. Generation of synthetic training data for SEEG electrodes segmentation. Int. J. Comput. Assist. Radiol. Surg. 17 , 937–943 (2022). Su, D. et al. Automatic, deep-learning-based segmentation of the amygdalohippocampectomy resection cavity in MR images. in Medical Imaging 2023: Image-Guided Procedures, Robotic Interventions, and Modeling (eds. Linte, C. A. & Siewerdsen, J. H.) 33 (SPIE, 2023). Courtney, M. R. et al. Automated segmentation of epilepsy surgical resection cavities: Comparison of four methods to manual segmentation. Neuroimage 296 , 120682 (2024). Pantovic, A., Ollivier, I. & Essert, C. 2D and 3D-UNet for segmentation of SEEG electrode contacts on post-operative CT scans. in Medical Imaging 2022: Image-Guided Procedures, Robotic Interventions, and Modeling (eds. Linte, C. A. & Siewerdsen, J. H.) (SPIE, 2022). doi:10.1117/12.2606538. Banerjee, S. et al. Deep relational reasoning for the prediction of language impairment and postoperative seizure outcome using preoperative DWI connectome data of children with focal epilepsy. IEEE Trans. Med. Imaging 40 , 793–804 (2021). Zhang, W. et al. Matpr-Unet: A Multi Attention Two-Path Residual Unet for Focal Cortical Dysplasia Lesions Segmentation. in ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) 1641–1645 (IEEE, 2024). Johnson, G. W. et al. Localizing seizure onset zones in surgical epilepsy with neurostimulation deep learning. Journal of neurosurgery 138 , 1002–1007 (2022). Fıçıcı, C., Telatar, Z., Koçak, O. & Eroğul, O. Identification of TLE focus from EEG signals by using deep learning approach. Diagnostics (Basel) 13 , (2023). Sharma, R. Localization of epileptic surgical area using automated hybrid approach based on higher-order statistics with sensitivity analysis and residual wavelet transform. Biomed. Signal Process. Control 86 , 105192 (2023). Wang, X., Wu, D. & Yang, C. Localization of epileptic foci from intracranial EEG using the GRU-GC algorithm. Brain Inform. 12 , 6 (2025). Zhang, Z., Luo, Q. & Zhou, Y. Localization of focal intractable epileptic focus based on long-term recurrent convolutional network. in 2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI) 570–574 (IEEE, 2023). Wang, Y. et al. SEEG-Net: An explainable and deep learning-based cross-subject pathological activity detection method for drug-resistant epilepsy. Comput. Biol. Med. 148 , 105703 (2022). Dou, Y. et al. Identification of epileptic networks with graph convolutional network incorporating oscillatory activities and evoked synaptic responses. Neuroimage 284 , 120439 (2023). Chen, W. et al. Unsupervised detection of high-frequency oscillations in intracranial electroencephalogram: promoting a valuable automated diagnostic tool for epilepsy. Front. Neurol. 16 , 1455613 (2025). Norris, J. et al. Localising the Seizure Onset Zone from Single-Pulse Electrical Stimulation Responses with a CNN Transformer. in Proceedings of the 9th Machine Learning for Healthcare Conference: PMLR (eds. Deshpande, K. et al.) vol. 252 1–21 (PMLR, Toronto, Canada, 2024). Li, S., Wang, Y., Yang, Y., Wei, P. & Kang, G. CRM: An automatic label generation method based on semi-supervised learning for high frequency oscillatory. in Proceedings of the 2022 9th International Conference on Biomedical and Bioinformatics Engineering 100–107 (ACM, New York, NY, USA, 2022). Liu, X., Hu, L., Wang, S. & Shen, J. Localization of seizure onset zone with epilepsy propagation networks based on graph convolutional network. Biomed. Signal Process. Control 74 , 103489 (2022). Monsoor, T. et al. Optimizing detection and deep learning-based classification of pathological high-frequency oscillations in epilepsy. Clin. Neurophysiol. 154 , 129–140 (2023). Sadek, Z., Hadriche, A. & Jmail, N. Clustering of high frequency oscillations HFO in epilepsy using pretrained neural networks. in Lecture Notes in Networks and Systems 100–107 (Springer Nature Switzerland, Cham, 2023). Liu, X. et al. Focal EEG recognition combining LMD and deep reinforcement learning. in Proceedings of the 2023 3rd International Conference on Bioinformatics and Intelligent Computing 165–168 (ACM, New York, NY, USA, 2023). Gardy, L. et al. Detecting fast-ripples on both micro- and macro-electrodes in epilepsy: A wavelet-based CNN detector. J. Neurosci. Methods 415 , 110350 (2025). Broti, N. M. et al. Multi-modal multitask learning model for simultaneous classification of two epilepsy biomarkers. in Lecture Notes in Computer Science 235–250 (Springer Nature Switzerland, Cham, 2025). Additional Declarations No competing interests reported. Supplementary Files AIenableddecisionsupportsystemsinepilepsysurgeryascopingreviewsupplementary.pdf Cite Share Download PDF Status: Under Review Version 1 posted Editorial decision: Revision requested 11 Mar, 2026 Reviews received at journal 08 Mar, 2026 Reviewers agreed at journal 26 Feb, 2026 Reviewers agreed at journal 31 Jan, 2026 Reviews received at journal 30 Jan, 2026 Reviewers agreed at journal 26 Jan, 2026 Reviewers invited by journal 26 Jan, 2026 Editor assigned by journal 20 Jan, 2026 Submission checks completed at journal 19 Jan, 2026 First submitted to journal 15 Jan, 2026 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-8612799","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Article","associatedPublications":[],"authors":[{"id":591899397,"identity":"5c33182d-8e41-4b61-92b3-2e4ffa917325","order_by":0,"name":"Kai Yu","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Kai","middleName":"","lastName":"Yu","suffix":""},{"id":591899398,"identity":"de380b50-225d-4bc8-ad7e-e1057ff2a816","order_by":1,"name":"Shuang Zhou","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Shuang","middleName":"","lastName":"Zhou","suffix":""},{"id":591899399,"identity":"a911c95f-e58c-4bba-b7df-b1dc34873b35","order_by":2,"name":"Meijia Song","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Meijia","middleName":"","lastName":"Song","suffix":""},{"id":591899400,"identity":"b68784a4-9135-444e-81c8-d21b0a134bd4","order_by":3,"name":"Zaifu Zhan","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Zaifu","middleName":"","lastName":"Zhan","suffix":""},{"id":591899401,"identity":"6fc172ca-a5d8-4c34-beeb-8c7ce7cb4f61","order_by":4,"name":"Yu Hou","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Yu","middleName":"","lastName":"Hou","suffix":""},{"id":591899402,"identity":"ad5895e3-d6bb-420b-8b26-88433a6ff662","order_by":5,"name":"Yiran Song","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Yiran","middleName":"","lastName":"Song","suffix":""},{"id":591899403,"identity":"c7f78341-59df-49e2-ab70-587062b0209a","order_by":6,"name":"Min Zeng","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Min","middleName":"","lastName":"Zeng","suffix":""},{"id":591899404,"identity":"061a1aea-6f17-4cda-9019-94fba0585706","order_by":7,"name":"Biao Yin","email":"","orcid":"","institution":"University of Massachusetts Chan Medical School","correspondingAuthor":false,"prefix":"","firstName":"Biao","middleName":"","lastName":"Yin","suffix":""},{"id":591899405,"identity":"2c369478-0fc1-4ebd-a90f-34936ea05ac4","order_by":8,"name":"Feifan Liu","email":"","orcid":"","institution":"University of Massachusetts Chan Medical School","correspondingAuthor":false,"prefix":"","firstName":"Feifan","middleName":"","lastName":"Liu","suffix":""},{"id":591899406,"identity":"3b0ad781-fe3d-40b2-a247-1089c8dac7ad","order_by":9,"name":"Sandipan Pati","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Sandipan","middleName":"","lastName":"Pati","suffix":""},{"id":591899407,"identity":"f79df299-9b27-46f9-839c-f53dfd2b3a0d","order_by":10,"name":"Zhiyi Sha","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Zhiyi","middleName":"","lastName":"Sha","suffix":""},{"id":591899408,"identity":"d9689819-18ec-49f4-b325-5da93db14e6e","order_by":11,"name":"Mingquan Lin","email":"","orcid":"","institution":"University of Minnesota","correspondingAuthor":false,"prefix":"","firstName":"Mingquan","middleName":"","lastName":"Lin","suffix":""},{"id":591899409,"identity":"b3cb76e0-60eb-493b-9789-03c9b345dbbe","order_by":12,"name":"Rui Zhang","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAAuUlEQVRIiWNgGAWjYFACxgZmEMUP4TGToEWygXgtUGUGB4jVYj4juflzQc0du803stMeMFRYJzYQ0iJzI7HBeMaxZ8nbbuRuN2A4k05Yi4REYkMyD9vhZLMbudskGNsOE6flMM+/w8nGM0Ba/hGnpbGZt+2wnYEESEsDMVp4HjYz8/YdTpA483abRMKxdGPCWtjTH3/m+XbYnr8daMuHGmtZglpgAOKeBGKVg4A9KYpHwSgYBaNghAEANgI+i397l8IAAAAASUVORK5CYII=","orcid":"","institution":"University of Minnesota","correspondingAuthor":true,"prefix":"","firstName":"Rui","middleName":"","lastName":"Zhang","suffix":""}],"badges":[],"createdAt":"2026-01-15 17:38:47","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-8612799/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-8612799/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":102982270,"identity":"dd032b17-5e72-416d-be5b-bf5d3cffd78b","added_by":"auto","created_at":"2026-02-19 09:14:03","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":908897,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eOverview of AI-enabled decision support systems in epilepsy surgery and the key analytical dimensions of this scoping review. \u003c/strong\u003eThe figure summarizes how the included studies were organized across the epilepsy surgery pathway, including surgical stages and clinical tasks, dataset characteristics and data modalities, modeling approaches and training strategies, and levels of workflow integration, together with publication trends used to synthesize the evidence base. Icons sourced from Flaticon.com (full attributions in Supplementary Note 1).\u003c/p\u003e","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/ac5208cbbb4d5797724b7a58.png"},{"id":103049729,"identity":"ccdd6cd7-6981-4d84-a9ed-c941cd33e110","added_by":"auto","created_at":"2026-02-20 07:45:14","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":431943,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSankey diagram of AI-enabled decision support studies in epilepsy surgery.\u003c/strong\u003e Flows map included studies from surgical stage to target clinical task, data modality, and model architecture. Link width is proportional to the number of studies along each pathway.\u003c/p\u003e","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/ef685ed9de8e4aabd8368d97.png"},{"id":103049929,"identity":"e23a92d1-6af3-4bf5-a740-49d94557f9af","added_by":"auto","created_at":"2026-02-20 07:47:17","extension":"png","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":199407,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eClinical contexts and dataset characteristics of included studies in epilepsy surgery.\u003c/strong\u003e (a) Targeted clinical tasks. (b) Surgical stage coverage (pre-operative, post-operative, or both). (c) Data modalities used. (d) Geographic provenance of datasets. (e) Cumulative distribution of sample sizes across studies (log scale). (f) Dataset accessibility (private, public, or mixed).\u003c/p\u003e","description":"","filename":"floatimage3.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/3eaccb3e531ffda5b50f24cf.png"},{"id":102982271,"identity":"51f8c53d-3044-4105-a29f-04011e761e1f","added_by":"auto","created_at":"2026-02-19 09:14:04","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":163660,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eMethodological and implementation-relevant characteristics of included studies in epilepsy surgery.\u003c/strong\u003e (a) Model architecture categories. (b) External validation (none, single-center, or multi-center). (c) Training strategies. (d) Evaluation approach (automatic, manual, or mixed). (e) Workflow integration stages (offline research, decision-support tool, or integrated system).\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/b0dd1c8687e453f0315df035.png"},{"id":102982277,"identity":"2d1bb13e-6307-4c11-bf7a-3520a7bfcc2f","added_by":"auto","created_at":"2026-02-19 09:14:04","extension":"png","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":153894,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePublication landscape of included AI-enabled decision support studies in epilepsy surgery (2018-2025).\u003c/strong\u003e(a) Publication types (journal article, or conference paper). (b) Annual publication trends by region. (c) Regional distribution. (d) Country distribution.\u003c/p\u003e","description":"","filename":"floatimage5.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/457a3c3a6b362a1f8b9e1e46.png"},{"id":102982274,"identity":"9f256fef-337c-49f3-a041-91a316c7e5b9","added_by":"auto","created_at":"2026-02-19 09:14:04","extension":"png","order_by":6,"title":"Figure 6","display":"","copyAsset":false,"role":"figure","size":1008501,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSummary of limitations, future directions, and key take-home messages for AI-enabled decision support systems in epilepsy surgery.\u003c/strong\u003e The figure summarizes major gaps in pathway coverage, data resources, modeling and validation, evaluation and translation, and global participation, and highlights corresponding research priorities to support safe, scalable, and equitable adoption.\u003c/p\u003e","description":"","filename":"floatimage6.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/68036f4e5d6ab7ebe3eb8e54.png"},{"id":102982275,"identity":"f01fb654-5db2-4e20-8bf6-75795835a0bc","added_by":"auto","created_at":"2026-02-19 09:14:04","extension":"png","order_by":7,"title":"Figure 7","display":"","copyAsset":false,"role":"figure","size":349490,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003ePRISMA-ScR flowchart of study selection process.\u003c/strong\u003e The flowchart summarizes the numbers of records identified from databases, duplicates removed, records screened, full-text articles assessed for eligibility, studies included, and the main reasons for exclusion at each stage.\u003c/p\u003e","description":"","filename":"floatimage7.png","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/d1ad66148e7e4bb2c7a7f141.png"},{"id":103051150,"identity":"d37bd75b-50d3-45a8-b9ac-2d40e8f086eb","added_by":"auto","created_at":"2026-02-20 07:58:36","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":4131921,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/40b8ee55-8d58-4fc6-870f-719618c2d67b.pdf"},{"id":103049884,"identity":"f2d6ec83-a9aa-4e5e-b13c-6191822d0a2d","added_by":"auto","created_at":"2026-02-20 07:47:02","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"supplement","size":510836,"visible":true,"origin":"","legend":"","description":"","filename":"AIenableddecisionsupportsystemsinepilepsysurgeryascopingreviewsupplementary.pdf","url":"https://assets-eu.researchsquare.com/files/rs-8612799/v1/bca1f5b77bf2daf059a916c4.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"AI enabled decision support systems in epilepsy surgery a scoping review","fulltext":[{"header":"Introduction","content":"\u003cp\u003eEpilepsy surgery is a high-risk, resource-intensive, multidisciplinary care pathway that requires coordinated, time-sensitive decisions across pre-operative, intra-operative, and post-operative stages. Epilepsy affects more than 50\u0026nbsp;million people worldwide\u003csup\u003e\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e\u003c/sup\u003e, and global epidemiological estimates suggest that approximately 10.1\u0026nbsp;million people living with epilepsy are potential surgical candidates, with around 1.4\u0026nbsp;million new surgically treatable cases each year\u003csup\u003e\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e\u003c/sup\u003e. Despite its promise for drug-resistant epilepsy, surgical success depends on integrated presurgical evaluation, precise operative execution, and careful post-operative monitoring to maximize seizure control while minimizing the risk of neurological complications\u003csup\u003e\u003cspan additionalcitationids=\"CR4\" citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e\u003c/sup\u003e. Because workflows, expertise, and data infrastructures vary across centers, achieving consistent decision-making and scalable implementation remains a health-system challenge.\u003c/p\u003e \u003cp\u003eAcross the pre-operative, intra-operative, and post-operative phases of epilepsy surgery, clinicians rely on a wide range of diagnostic modalities, including magnetic resonance imaging (MRI), positron emission tomography (PET), computed tomography (CT), scalp electroencephalography (EEG), and intracranial EEG, to support multidisciplinary decisions such as lesion detection, seizure-onset (SOZ) and epileptogenic zone (EZ) localization, electrode navigation, and postoperative outcome assessment. However, these decisions rarely rely on any single modality. Instead, they depend on how multimodal data are synthesized and interpreted, a process that remains inherently complex and variable. Interpretation is often influenced by center-specific practices, and clinician experience, leading to substantial inter-institution variability even when similar diagnostic modalities are available. This variability complicates standardization and limits the scalability of consistent decision support across epilepsy centers.\u003c/p\u003e \u003cp\u003eBeyond interpretive subjectivity, epilepsy surgery workflows are further challenged by the substantial specialist time and expertise required to interrogate complex diagnostic data. Depending on the clinical question, clinicians may need to review hours to weeks of scalp or intracranial EEG recordings, scrutinize high-resolution structural MRI for subtle cortical abnormalities, interpret regional metabolic patterns on PET, or evaluate post-operative structural changes on CT. These datasets are high-dimensional, noisy, acquired under heterogeneous protocols, and demand specialized expertise that varies across clinicians and centers. As a result, manual analysis is not only time-intensive but also vulnerable to inter-observer variability, limiting standardization and scalability in surgical decision-making. These constraints motivate workflow-integrated, scalable decision-support systems that can distill clinically meaningful features and enable more reproducible evaluation across all stages of epilepsy surgery.\u003c/p\u003e \u003cp\u003eWith the rapid development of artificial intelligence (AI), an increasing number of studies have explored AI-enabled decision support across the epilepsy surgery pathway. Most of this literature is based on deep learning methods. These approaches have shown promise in tasks such as structural MRI\u0026ndash;based lesion detection\u003csup\u003e\u003cspan additionalcitationids=\"CR7\" citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e\u003c/sup\u003e, intracranial EEG\u0026ndash;based SOZ/EZ localization\u003csup\u003e\u003cspan additionalcitationids=\"CR10\" citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e11\u003c/span\u003e\u003c/sup\u003e, automated functional mapping\u003csup\u003e\u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e,\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u003c/sup\u003e, surgical planning and navigation\u003csup\u003e\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e\u003c/sup\u003e, and surgical outcomes prediction\u003csup\u003e\u003cspan additionalcitationids=\"CR18 CR19\" citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e\u003c/sup\u003e. A small number of studies have also begun to investigate foundation model approaches, for example large language models (LLMs) for clinical text interpretation\u003csup\u003e\u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e\u003c/sup\u003e. However, the literature remains largely focused on technical feasibility in controlled settings. Many systems remain offline research prototypes, some are developed as standalone decision-support tools\u003csup\u003e\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e,\u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e\u003c/sup\u003e, and a small number have been partially integrated into existing clinical workflows\u003csup\u003e\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e\u003c/sup\u003e. Health systems still lack the evidence needed to decide whether and how these tools can be deployed. Cross-center validation and evaluations of workflow integration, patient impact, and safety are still uncommon.\u003c/p\u003e \u003cp\u003eCurrent research on AI-enabled decision support in epilepsy surgery remains fragmented across clinical tasks, data modalities, and study designs, making it difficult to judge where evidence is mature enough for implementation across the full care pathway. Existing review articles have largely reflected this fragmentation. Many focus on a single component of the surgical workflow or on a single modality, which limits the ability of clinicians and health systems to compare approaches, prioritize prospective evaluation, and plan workflow integration. For example, several reviews concentrate on EEG or intracranial EEG analysis\u003csup\u003e\u003cspan additionalcitationids=\"CR26 CR27 CR28 CR29\" citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e\u003c/sup\u003e, providing insights into electrophysiology-based SOZ/EZ localization but offering limited discussion of imaging-based approaches or post-operative evaluation. Conversely, neuroimaging-focused reviews centered on MRI, PET, or CT often exclude electrophysiological studies entirely, resulting in an incomplete and modality-isolated picture of current evidence\u003csup\u003e\u003cspan additionalcitationids=\"CR32\" citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e\u003c/sup\u003e. This separation across modalities and tasks prevents a pathway level assessment of AI systems, including how they are validated, evaluated for clinical impact, and moved toward workflow integration. Overall, a unified synthesis through a health-systems lens is needed to map the distribution of evidence, characterize methodological choices, and identify translational gaps that constrain real-world adoption.\u003c/p\u003e \u003cp\u003eTo address these gaps, we focus on five guiding questions: \u003cb\u003e(RQ1)\u003c/b\u003e How are AI approaches distributed across the surgical pathway, including their targeted clinical tasks and the data modalities used? \u003cb\u003e(RQ2)\u003c/b\u003e What are the characteristics of the datasets underpinning these studies, including their geographic origins, sample sizes, and accessibility? \u003cb\u003e(RQ3)\u003c/b\u003e What modeling approaches and training strategies have been adopted, and what evidence is reported to support generalizability, including external validation? \u003cb\u003e(RQ4)\u003c/b\u003e How is model performance assessed and reported, and how far have these systems advanced toward workflow integration and real-world deployment? \u003cb\u003e(RQ5)\u003c/b\u003e How have research trends evolved over time, including global participation and institutional contributions, and what do they suggest about the maturity of the field?\u003c/p\u003e \u003cp\u003eTo answer these questions, this scoping review provides a comprehensive and structured examination of AI-enabled decision support, primarily deep learning-based systems, across the epilepsy surgery pathway. We first summarize the clinical context of existing studies, including their surgical stages, targeted clinical tasks, and data modalities. We then characterize the underlying datasets by assessing their geographic sources, sample sizes, and accessibility. Next, we synthesize modeling approaches and training strategies, with particular attention to evidence supporting generalizability, including external validation practices. We also examine how performance is assessed and reported, and how far systems have advanced toward workflow integration and real-world deployment. Finally, we review publication patterns and institutional contributions to contextualize the evolution of the field. Together, these analyses map the current landscape, clarify implementation-relevant evidence gaps, and highlight priorities for developing and evaluating AI systems that can be adopted in real-world care pathways and across centers.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"No\" id=\"Taba\" border=\"1\"\u003e \u003ccolgroup cols=\"1\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eBox 1 | Terms and Concepts\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eLarge Language Models (LLMs)\u003c/b\u003e: Large-scale neural networks trained on massive text corpora to acquire general language understanding and reasoning capabilities. LLMs can interpret clinical narratives, summarize reports, assist decision-making, and serve as generalizable backbones for medical tasks.\u003c/p\u003e \u003cp\u003e\u003cb\u003eLLM Prompting\u003c/b\u003e: Providing tailored instructions, demonstrations, or constraints to an LLM to guide its output toward a target task without fine-tuning. Includes zero-shot, few-shot, and chain-of-thought.\u003c/p\u003e \u003cp\u003e\u003cb\u003eRetrieval-Augmented Generation (RAG)\u003c/b\u003e: A framework that combines external knowledge retrieval with generative modeling, enabling models to produce responses that are accurate, up-to-date, and grounded in verifiable sources.\u003c/p\u003e \u003cp\u003e\u003cb\u003eFoundation Models (FMs)\u003c/b\u003e: Large-scale pre-trained models that learn generalizable representations across modalities or tasks, providing adaptable backbones for downstream medical applications.\u003c/p\u003e \u003cp\u003e\u003cb\u003eTransfer Learning\u003c/b\u003e: Fine-tuning models pre-trained on large datasets (e.g., ImageNet MRI models) for epilepsy-specific tasks.\u003c/p\u003e \u003cp\u003e\u003cb\u003eSelf-supervised Learning (SSL)\u003c/b\u003e: Training models using proxy tasks (e.g., contrastive learning, masked modeling) to learn robust representations from unlabeled neuroimaging or EEG data.\u003c/p\u003e \u003cp\u003e\u003cb\u003eSemi-supervised Learning\u003c/b\u003e: Training with both labeled and unlabeled data, often using pseudo-labeling or consistency regularization to mitigate label scarcity.\u003c/p\u003e \u003cp\u003e\u003cb\u003eUnsupervised Learning\u003c/b\u003e: Learning patterns without explicit labels\u0026mdash;for example anomaly detection, clustering, or autoencoder feature representation.\u003c/p\u003e \u003cp\u003e\u003cb\u003eReinforcement Learning (RL)\u003c/b\u003e: Learning decision policies from reward signals, occasionally applied to trajectory planning or electrode optimization.\u003c/p\u003e \u003cp\u003e\u003cb\u003eFusion Strategies\u003c/b\u003e: Definitions of early, intermediate, late, and hybrid fusion, describing how multimodal data are integrated.\u003c/p\u003e \u003cp\u003e\u003cb\u003eExternal Validation\u003c/b\u003e: Testing model generalizability on independent datasets from different centers or regions, categorized as single-center or multi-center external validation.\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e"},{"header":"Results","content":"\u003ch2\u003eOverview of the Scope\u003c/h2\u003e\n\u003cp\u003eThis section provides an overview of the scoping review and summarizes its principal findings (Fig. 1). We synthesize the literature on AI-enabled decision support in epilepsy surgery across five implementation-relevant dimensions: clinical context along the surgical pathway, dataset characteristics, modeling approaches and evidence supporting generalizability, clinical integration, and research trends.\u003c/p\u003e\n\u003cp\u003eAs illustrated in Fig. 2 and Fig. 3, we first map AI applications across the surgical pathway by characterizing targeted clinical tasks and associated data modalities \u003cstrong\u003e(RQ1)\u003c/strong\u003e. We then examine the datasets used in these studies (Fig. 3), including their geographic origins, sample sizes, and levels of accessibility \u003cstrong\u003e(RQ2)\u003c/strong\u003e. Next, we synthesize modeling approaches and training strategies (Fig. 4), with attention to evidence reported to support generalizability, including external validation \u003cstrong\u003e(RQ3)\u003c/strong\u003e. We also review how performance is assessed and reported, and how far systems have advanced toward workflow integration and real-world deployment \u003cstrong\u003e(\u003c/strong\u003eFig. 4; \u003cstrong\u003eRQ4)\u003c/strong\u003e. Finally, in Fig. 5, we contextualize the evolution of the field by summarizing publication trends, study types, and regional contributions \u003cstrong\u003e(RQ5)\u003c/strong\u003e. Collectively, these results provide a pathway-level view of where evidence is concentrated and where implementation-relevant gaps remain, helping to prioritize future prospective evaluation and workflow integration efforts for adoption across centers.\u003c/p\u003e\n\u003ch2\u003eStudy Selection and Characteristics\u003c/h2\u003e\n\u003cp\u003eFig. 7 summarizes the study selection process. Of the 2,534 records initially identified, 1,299 duplicates were removed, leaving 1,235 records for title and abstract screening. At this stage, 854 records were excluded for not meeting the predefined inclusion criteria, resulting in 381 reports eligible for full-text review. During full-text screening, 236 records were further excluded due to the following reasons: out-of-scope setting or context (N=5), ineligible study design (N=54), or irrelevant methods (N=177). Ultimately, 145 studies were included in this scoping review.\u003c/p\u003e\n\u003cp\u003eAcross the included studies, evidence was heavily concentrated in the pre-operative stage (91%, N=132), with relatively few post-operative studies (8.3%, N=12), and one study spanning both stages. No included studies specifically addressed the intra-operative stage, highlighting a clear gap in pathway coverage. Within the pre-operative stage, the most common tasks were SOZ/EZ localization (60.7%, N=88), and lesion detection/classification (18.6%, N=27), followed by presurgical functional mapping (6.2%, N=9) and surgical planning and navigation (2.1%, N=3). Post-operative studies primarily focused on post-operative assessment and outcome prediction (7.6%, N=11), with one study addressing electrode localization/navigation support (0.7%, N=1). Data modalities varied by task, with electrophysiology data (scalp EEG, intracranial EEG, and MEG) commonly used for SOZ/EZ localization, and structural MRI predominantly used for lesion detection and classification. Dataset provenance showed broad geographic coverage, with 29.7% (N=43) using data from North America, 26.9% (N=39) from Asia, 26.9% (N=39) from Europe, and 3.5% (N=5) from Oceania. Multi-regional datasets were used in 12.4% of studies (N=18), reflecting an emerging trend toward broader data diversity. Modeling approaches were dominated by CNN-based methods (46.2%, N=67), followed by U-Net/FCN-based segmentation models (14.5%, N=21), with smaller proportions of hybrid CNN\u0026ndash;RNN models (8.3%, N=12) and GNN-based architectures (5.5%, N=8). Two studies (1.4%) used large language models (LLMs), specifically ChatGPT-4, reflecting early exploration of foundation model approaches in this domain. Additional LLM-based studies\u003csup\u003e34,35\u003c/sup\u003e identified during screening were excluded because they focused on surgical candidate selection or pre-surgical triage rather than tasks within the epilepsy surgery pathway, consistent with our predefined criteria. Implementation readiness remained limited. Most studies were conducted as offline research (89.6%, N=130), a smaller proportion functioned as decision-support tools (7.6%, N=11), and only four studies (2.8%) were integrated into clinical systems. Publication output consisted of journal articles (74.5%, N=108) and conference papers (25.5%, N=37), with activity increasing over time, including notable increases in 2020 (16.6%, N=24) and 2022 (21.4%, N=31).\u003c/p\u003e\n\u003ch2\u003eClinical Context\u003c/h2\u003e\n\u003cp\u003eAs illustrated in Fig. 3a-c, we summarize the clinical contexts of the included studies, including their targeted clinical tasks, surgical stages, and data modalities. Among these studies, 91% (N=132) focused on the pre-operative stage, with the main clinical tasks involving SOZ/EZ localization, lesion detection/classification, presurgical functional mapping, and surgical planning and navigation. Notably, 60.7% (N=88) of all included studies centered on SOZ/EZ localization, making it the predominant research focus in current deep learning applications for epilepsy surgery. For example, recent studies have leveraged various electrophysiology modalities, including scalp EEG\u003csup\u003e23,36\u0026ndash;43\u003c/sup\u003e, intracranial EEG\u003csup\u003e9\u0026ndash;11,44\u0026ndash;69\u003c/sup\u003e, and MEG\u003csup\u003e70\u0026ndash;77\u003c/sup\u003e, to identify seizure-onset and epileptogenic zones. Other works have utilized MRI based data, such as structural MRI\u003csup\u003e78\u0026ndash;80\u003c/sup\u003e and resting-state fMRI\u003csup\u003e81\u0026ndash;85\u003c/sup\u003e, and two studies have analyzed clinical text\u003csup\u003e21,86\u003c/sup\u003e for SOZ/EZ localization. Beyond single modalities, several works also incorporated multimodal inputs. For instance, Li et al.\u003csup\u003e87\u003c/sup\u003e and Yang et al.\u003csup\u003e88\u003c/sup\u003e combined intracranial EEG and scalp EEG for SOZ/EZ localization. In addition, multiple studies such as Jeong et al.\u003csup\u003e89\u003c/sup\u003e, Fard et al.\u003csup\u003e90\u003c/sup\u003e, Banerjee et al.\u003csup\u003e91\u003c/sup\u003e, and Zotova et al.\u003csup\u003e92\u003c/sup\u003e employed various imaging modalities to perform SOZ/EZ localization, including MRI (T1, T2, FLAIR, and DWI) and PET. A total of 27 studies (18.6%) focused on lesion detection/classification task, most of which relied on structural MRI with multiple sequences\u003csup\u003e6\u0026ndash;8,24,93\u0026ndash;108\u003c/sup\u003e, Intracranial EEG\u003csup\u003e109,110\u003c/sup\u003e, pathology\u003csup\u003e111\u003c/sup\u003e, or multimodal inputs\u003csup\u003e112\u0026ndash;114\u003c/sup\u003e for lesion detection/classification. Nine studies (6.2%) focused on presurgical functional mapping, and commonly used data modalities included diffusion MRI tractography\u003csup\u003e13,115\u0026ndash;117\u003c/sup\u003e, motion visual analysis\u003csup\u003e12,118,119\u003c/sup\u003e, intracranial EEG\u003csup\u003e120\u003c/sup\u003e, and audio speech\u003csup\u003e121\u003c/sup\u003e. The remaining three studies (2.1%) focused on surgical planning and navigation. Granados et al.\u003csup\u003e14\u003c/sup\u003e proposed an early-fusion multimodal approach integrating T1-weighted MRI, DWI, and CT. Nejedly et al.\u003csup\u003e16\u003c/sup\u003e introduced a late-fusion multimodal method based on intracranial EEG and T1-weighted MRI. In contrast, Liu et al.\u003csup\u003e15\u003c/sup\u003e relied solely on T1-weighted MRI to perform surgical planning and navigation. Unlike the single-task studies summarized above, five studies addressed multiple pre-operative tasks. Hossain et al.\u003csup\u003e122\u003c/sup\u003e and Zhang et al.\u003csup\u003e123\u003c/sup\u003e used scalp EEG and intracranial EEG, respectively, to perform both SOZ/EZ localization and presurgical functional mapping. Mo et al.\u003csup\u003e124\u003c/sup\u003e employed an early-fusion multimodal approach integrating MRI sequences with T1, FLAIR and PET to achieve lesion detection/classification and surgical planning and navigation. Similarly, Park et al.\u003csup\u003e125\u003c/sup\u003e applied an early-fusion multimodal strategy, based on T1-weighted MRI and T2-FLAIR, to perform SOZ/EZ localization and lesion detection/classification. In addition, Hu et al.\u003csup\u003e126\u003c/sup\u003e used scalp EEG for both SOZ/EZ localization and surgical planning and navigation.\u003c/p\u003e\n\u003cp\u003eIn contrast, 8.3% (N=12) of studies addressed the post-operative stage, primarily targeting post-operative assessment and outcome prediction\u003csup\u003e17\u0026ndash;20,22,127\u0026ndash;132\u003c/sup\u003e and electrode localization/navigation support\u003csup\u003e133\u003c/sup\u003e. A variety of data modalities were applied in this stage, including diffusion tractography MRI\u003csup\u003e20,127\u003c/sup\u003e, structural MRI\u003csup\u003e22,128,129,131,132\u003c/sup\u003e, CT\u003csup\u003e130,133\u003c/sup\u003e, and electrophysiology (Intracranial EEG)\u003csup\u003e17,19\u003c/sup\u003e. Notably, Tang et al.\u003csup\u003e18\u003c/sup\u003e employed a multimodal late-fusion approach that integrated multi-sequence MRI (T1, T2, and FLAIR), CT, and PET to achieve post-operative assessment and outcome prediction. Additionally, one study examined both stages\u003csup\u003e134\u003c/sup\u003e, covering presurgical functional mapping in the pre-operative phase and post-operative assessment and outcome prediction in the post-operative phase.\u003c/p\u003e\n\u003cp\u003eOverall, current AI-enabled decision-support research in epilepsy surgery remains heavily concentrated in the pre-operative stage, with SOZ/EZ localization and structural MRI\u0026ndash;based lesion detection/classification accounting for most studies. In contrast, presurgical functional mapping, surgical planning and navigation, and post-operative evaluation remain relatively underrepresented, and no study in our review specifically addressed intra-operative applications. Although multimodal fusion has begun to appear in recent work, its adoption is still limited, and comprehensive multi-task models remain uncommon. The absence of intra-operative studies also indicates limited evidence for real-time, workflow-integrated decision support during surgery.\u003c/p\u003e\n\u003ch2\u003eData Characteristics\u003c/h2\u003e\n\u003cp\u003eAcross the included studies, 10 different data modalities and types were used. Electrophysiology was the most common (52.8%, N=84), primarily consisting of intracranial EEG\u003csup\u003e9\u0026ndash;11,44\u0026ndash;69\u003c/sup\u003e, scalp EEG\u003csup\u003e23,36\u0026ndash;43\u003c/sup\u003e, and MEG\u003csup\u003e70\u0026ndash;77\u003c/sup\u003e. This was followed by structural MRI (25.2%, N=40), mainly including T1-weighted, T2-weighted, and FLAIR MRI; diffusion MRI (6.9%, N=11), including DWI\u003csup\u003e13,14,20,89,91,115,116,134\u003c/sup\u003e, DTI\u003csup\u003e127\u003c/sup\u003e, and DKI\u003csup\u003e108\u003c/sup\u003e; PET\u003csup\u003e90,124\u003c/sup\u003e (4.4%, N=7); functional MRI (rs-fMRI)\u003csup\u003e81,84\u003c/sup\u003e (3.8%, N=6); and CT\u003csup\u003e133\u003c/sup\u003e (2.5%, N=4). One study use histopathology for FCD classification\u003csup\u003e111\u003c/sup\u003e. Seven studies incorporated video\u003csup\u003e12,118,119\u003c/sup\u003e, text\u003csup\u003e21,86\u003c/sup\u003e, or audio data\u003csup\u003e121\u003c/sup\u003e. Notably, many studies did not rely on a single modality for epilepsy surgery analysis. Instead, they employed combinations of multiple data modalities, such as the fusion of T2-weighted and FLAIR MRI\u003csup\u003e98,99\u003c/sup\u003e, intracranial EEG with rs-fMRI\u003csup\u003e112\u003c/sup\u003e, intracranial EEG with scalp EEG\u003csup\u003e88\u003c/sup\u003e, and the integration of multiple MRI sequences with PET\u003csup\u003e114,124,135\u003c/sup\u003e or CT\u003csup\u003e14,18\u003c/sup\u003e. Collectively, the breadth of modalities and heterogeneous acquisition protocols highlight the need for data harmonization and interoperable pipelines to support cross-center deployment.\u003c/p\u003e\n\u003cp\u003eBased on the geographic distribution of the datasets used in the included studies (Fig. 3d), most data originated from North America (29.7%, N=43), Asia (26.9%, N=39), and Europe (26.9%, N=39), with Oceania contributing 3.5% (N=5). In addition, 12.4% (N=18) of the studies utilized multi-regional datasets spanning two or more continents. For example, Yang et al.\u003csup\u003e88\u003c/sup\u003e combined intracranial EEG and scalp EEG data from Asia and Europe for pre-operative SOZ/EZ localization. Zhang et al.\u003csup\u003e135\u003c/sup\u003e similarly used data from Asia and Europe, but employed T1-weighted MRI, FLAIR MRI, and PET with a multimodal early-fusion strategy to achieve lesion detection/classification. Gill et al.\u003csup\u003e7\u003c/sup\u003e also adopted an early-fusion multimodal approach for lesion detection/classification, using T1-weighted MRI and FLAIR MRI collected from four regions: Asia, Europe, North America, and South America. The studies by Spitzer et al.\u003csup\u003e104\u003c/sup\u003e and Ripart et al.\u003csup\u003e106\u003c/sup\u003e incorporated the broadest geographic coverage to support lesion detection/classification, with T1-weighted and FLAIR MRI sourced from five regions (Asia, Europe, Oceania, North America, and South America). In contrast to the imaging- and electrophysiology-based studies above, Luo et al.\u003csup\u003e86\u003c/sup\u003e utilized clinical text collected from Asia, Europe, and North America to perform SOZ/EZ localization. Overall, the predominance of single-region datasets suggests that geographic representativeness and cross-center generalizability remain important constraints for implementation at scale.\u003c/p\u003e\n\u003cp\u003eMost studies relied on private datasets (65.5%, N=95), spanning diverse data types including audio, video, text, CT, PET, and multiple MRI sequences. In contrast, 22.1% (N=32) used fully public datasets, of which 30 involved electrophysiology data, including scalp EEG and intracranial EEG, and two studies used structural MRI (T1-weighted and FLAIR). Another 12.4% (N=18) employed mixed datasets combining private and public sources, mainly covering text, electrophysiology, PET, and T1-weighted MRI. The dataset sizes exhibited substantial heterogeneity, ranging from single-subject datasets to large cohorts exceeding 2,000 patients. The cumulative distribution (Fig. 3e) shows a strongly right-skewed, long-tailed pattern, with 80% of the studies using fewer than 100 patients. Only a small proportion of studies\u003csup\u003e40,78,87,95,111,135\u003c/sup\u003e used medium-sized datasets (100\u0026ndash;500 patients), and very few utilized large-scale cohorts\u003csup\u003e81,96,104\u003c/sup\u003e (\u0026gt;1,000 patients). Collectively, the dataset landscape is characterized by strong reliance on small, private, single-region datasets, with limited public data availability, which can impede external validation, reproducibility, and workflow-integrated evaluation needed for adoption across centers.\u003c/p\u003e\n\u003ch2\u003eModeling Approaches\u003c/h2\u003e\n\u003cp\u003eTo support AI-enabled decision support across diverse modalities, the included studies adopted a range of deep learning architectures (Fig. 4a). CNN-based models were the most common (46.2%, N=67), spanning 1D CNNs for electrophysiology\u003csup\u003e17,41,51,110,115,136\u003c/sup\u003e, 2D CNNs for MRI\u003csup\u003e84,85\u003c/sup\u003e, and 3D CNN for volumetric MRI\u003csup\u003e81,96,114,124\u003c/sup\u003e. In addition, 14.5% (N=21) used CNN-derived segmentation architectures, primarily U-Net or fully convolutional networks (FCNs), applied to PET\u003csup\u003e113,135\u003c/sup\u003e, CT\u003csup\u003e130,133\u003c/sup\u003e, and structural MRI data\u003csup\u003e8,15,22,95,98,100,102\u003c/sup\u003e. A further 8.3% (N=12) and 4.8% (N=7) of the studies adopted RNN-based\u003csup\u003e137\u0026ndash;139\u003c/sup\u003e and hybrid CNN\u0026ndash;RNN\u003csup\u003e140,141\u003c/sup\u003e architectures, respectively, most often for electrophysiology or video analysis. The remaining studies used a variety of model architectures, such as autoencoder-based\u003csup\u003e142,143\u003c/sup\u003e, Transformer-based\u003csup\u003e144\u003c/sup\u003e, and attention-based\u003csup\u003e68,70\u003c/sup\u003e models for specific tasks. Notably, one study\u003csup\u003e119\u003c/sup\u003e leveraged an existing detection framework (Mask R-CNN) to perform automated presurgical functional mapping from video data, and two studies\u003csup\u003e21,86\u003c/sup\u003e applied large language models (LLMs; ChatGPT-4) to clinical text for SOZ/EZ localization. For example, Luo et al.\u003csup\u003e86\u003c/sup\u003e addressed the SOZ/EZ localization task by using ChatGPT-4, which interpreted textual seizure semiology descriptions from literature and clinical records to automatically predict the most likely brain lobe of seizure onset.\u003c/p\u003e\n\u003cp\u003eThe training strategies were dominated by supervised learning, which accounted for 74.5% (N=108) of all models (Fig. 4c). Transfer learning was used in 8.3% (N=12) of the studies, typically through initialization with pretrained CNN backbones\u003csup\u003e18,37,118\u003c/sup\u003e. A smaller proportion employed semi-supervised learning\u003csup\u003e145\u0026ndash;147\u003c/sup\u003e (5.5%, N=8) or unsupervised learning techniques\u003csup\u003e143,148\u003c/sup\u003e (4.1%, N=6), such as autoencoder-based feature learning\u003csup\u003e143\u003c/sup\u003e. Only a limited number of studies utilized more advanced training paradigms, such as pretraining followed by finetuning\u003csup\u003e56,69,74,75,129\u003c/sup\u003e (3.4%, N=5), multi-training schemes\u003csup\u003e9,117\u003c/sup\u003e (1.4%, N=2), or reinforcement learning\u003csup\u003e149\u003c/sup\u003e (0.7%, N=1). A small number of studies required no additional training\u003csup\u003e119\u003c/sup\u003e (0.7%, N=1) or relied purely on LLM prompting\u003csup\u003e86\u003c/sup\u003e (0.7%, N=1). One study additionally incorporated retrieval-augmented generation (RAG) for text-based clinical reasoning\u003csup\u003e21\u003c/sup\u003e (0.7%, N=1). Overall, these patterns suggest that most systems depend on labeled data and task-specific training, which may limit scalability when high-quality annotations are difficult to obtain.\u003c/p\u003e\n\u003cp\u003eEvidence supporting generalizability remained limited (Fig. 4b). Most studies did not perform external validation (77.2%, N=112). A smaller subset conducted single-center external validation (17.2%, N=25), typically using an independent cohort from a different institution within the same geographic region. Several studies\u003csup\u003e23,57,65,86,88,104,135,145\u003c/sup\u003e validated models on external data from different geographic regions, providing a stronger test of cross-regional generalizability. Only 5.5% (N=8) of the studies\u003csup\u003e7,38,74,106,124,129,132,141\u003c/sup\u003e employed multi-center external validation, leveraging datasets from two or more independent sites to assess model generalizability across institutions. Taken together, while the field explores diverse architectures and training strategies, current epilepsy surgery AI is still dominated by supervised CNN-based approaches, and the scarcity of multi-center external validation indicates substantial gaps in implementation-relevant evidence for deployment across centers.\u003c/p\u003e\n\u003ch2\u003eEvaluation and Clinical Integration\u003c/h2\u003e\n\u003cp\u003eAs shown in Fig. 4d, most studies assessed model performance using fully automatic evaluation procedures (97.2%, N=141), in which quantitative performance metrics were computed directly from algorithmic outputs without additional human involvement. Common metrics included classification measures such as accuracy,\u0026nbsp;sensitivity, specificity, AUC, F1-score\u003csup\u003e40,78,84,144\u003c/sup\u003e, as well as segmentation measures such as IoU and DSC\u003csup\u003e7,106,135\u003c/sup\u003e. Only 1.4% (N=2) of the studies relied exclusively on manual evaluation. For example, Chiang et al.\u003csup\u003e21\u003c/sup\u003e developed an ontology-guided, GPT-based system to assist SOZ/EZ Localization from clinical text, and epilepsy specialists manually judged each model prediction as correct or incorrect to derive accuracy. Similarly, Pastore et al.\u003csup\u003e107\u003c/sup\u003e manually verified their algorithm generated lesion predictions against histopathology or scalp EEG findings. Another 1.4% (N=2) of studies employed mixed evaluation, combining automated quantitative metrics with expert review to adjudicate ambiguous cases or assess clinical plausibility. For instance, in Mo et al.\u003csup\u003e124\u003c/sup\u003e, sensitivity, specificity, and accuracy were automatically computed, but overlap validation and all trajectory feasibility and safety assessments were manually determined by clinical experts.\u003c/p\u003e\n\u003cp\u003eWith respect to clinical integration (Fig. 4e), most systems remained offline research prototypes (89.6%, N=130), without integration into routine clinical software environments or workflows. A smaller subset (7.6%, N=11) functioned as standalone decision-support tools\u003csup\u003e18,22,23,38,39,73,74,84,106,111,124\u003c/sup\u003e, in which clinicians accessed model outputs through independent research interfaces operating outside routine clinical systems such as PACS, SEEG analysis software, or platforms. In contrast, only 2.8% (N=4) of the studies reported semi-integrated clinical workflows, in which AI outputs were incorporated directly into interfaces already used in clinical practice, such as PACS viewers\u003csup\u003e107\u003c/sup\u003e, online SEEG review platforms\u003csup\u003e69\u003c/sup\u003e, or GUI-based electrophysiology analysis tools\u003csup\u003e150\u003c/sup\u003e, and prospectively validated MRI-based lesion detection systems applied in daily clinical reading\u003csup\u003e24\u003c/sup\u003e. These systems enabled clinicians to visualize, review, and interact with model predictions within existing workflows, but none were fully automated or embedded into hospital information systems.\u003c/p\u003e\n\u003cp\u003eTaken together, performance evaluation remains predominantly algorithm-centric and automated, while evidence for workflow-integrated use is limited. Most systems remain offline research tools, and only a small minority report partial, clinician-in-the-loop integration, highlighting ongoing gaps in implementation readiness for deployment across centers.\u003c/p\u003e\n\u003ch2\u003ePublication Trends\u003c/h2\u003e\n\u003cp\u003eThis section characterizes publication trends in epilepsy surgery AI research (Fig. 5). Overall, publication activity has increased substantially since 2018, with a notable rise between 2020 and 2022. Most studies were published as journals articles (74.5%, N=108), while conferences papers accounted for 25.5% (N=37) (Fig. 5a). Regionally, Asia and North America contributed the largest shares of publications (50.3%, N=73 and 32.4%, N=47, respectively). Followed by Europe (13.1%, N=19), Oceania (2.8%, N=4), and Africa (1.4%, N=2). At the country level, China produced the most studies (31.0%, N=45), followed by the United States (29.0%, N=42). India and Japan each contributed 7.6% (N=11), while France, the United Kingdom, and Canada contributed 4.8% (N=7), 3.4% (N=5), and 3.4% (N=5), respectively. The remaining countries collectively contributed 13.1% (N=19) of publications. Overall, these patterns indicate growing research activity and broadening international participation, while also highlighting uneven geographic distribution of evidence generation.\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eOur results highlight substantial gaps in pathway coverage and implementation readiness for AI-enabled decision support in epilepsy surgery. Evidence is heavily concentrated in the pre-operative stage (91%), with only 8.3% of studies addressing post-operative evaluation and none specifically targeting intra-operative applications. This imbalance reflects differences in data accessibility, operational constraints, and clinical risk. For example, pre-operative EEG and MRI are routinely acquired with established protocols, whereas intra-operative data collection is technically challenging, time-sensitive, and often constrained by surgical workflow. Beyond stage imbalance, effective seizure focus localization often requires integrating structural imaging, electrophysiology, and seizure semiology. However, most studies focus on either imaging (e.g., MRI) or electrophysiology (EEG) in isolation, and comparatively few integrate these modalities with the video-based behavioral and semiology information. Meanwhile, intra-operative AI remains substantially underdeveloped despite its clear clinical importance. Future research should prioritize evidence generation for intra-operative, workflow-integrated decision support\u003csup\u003e120\u003c/sup\u003e, and develop pathway-level approaches that connect pre-operative, intra-operative, and post-operative stages through end-to-end or multi-stage pipelines. Within pre-operative studies, SOZ/EZ localization and lesion detection/classification dominate current research activity. These tasks often have relatively well-defined reference standards (e.g., SEEG-confirmed seizure onset zones, postoperative seizure outcomes), making them easier to benchmark. In contrast, tasks such as presurgical functional mapping and surgical planning/navigation more closely reflect real clinical decision making but remain relatively underexplored. This imbalance in task coverage may constrain clinical impact by leaving critical components of the care pathway insufficiently supported. Although multimodal analyses can offer advantages over single-modality approaches, few studies conduct systematic comparisons of how different fusion strategies, including early\u003csup\u003e7,14,87,100,101\u003c/sup\u003e, late\u003csup\u003e16,18,112\u003c/sup\u003e, or intermediate\u003csup\u003e113,151\u003c/sup\u003e fusion, affect performance and interpretability. Advancing this field will require standardized and interoperable multimodal pipelines, along with interpretable cross-modal alignment methods that integrate electrophysiology, structural and functional imaging, and clinical information into unified decision-support systems.\u003c/p\u003e\n\u003cp\u003eGeneralizability and implementation readiness are constrained by the current data landscape. Across included studies, dataset size represents a major limitation. 80% of studies used small cohorts with fewer than 100 patients, while large-scale datasets (\u0026gt;1,000 patients) were extremely rare. This reflects the inherent difficulty of collecting large, high-quality epilepsy surgery datasets, particularly those involving multimodal imaging or invasive electrophysiology\u003csup\u003e23,92,126\u003c/sup\u003e. However, reliance on small datasets increases the risk of overfitting and limits the reliability of model performance estimates\u003csup\u003e12,44,76\u003c/sup\u003e. To address these challenges, future research should leverage advanced methods such as self-supervised learning, semi-supervised learning, and foundation models to enhance representation learning and improve generalization under small dataset conditions. Geographic and institutional coverage is similarly uneven. Most datasets originate from North America, Asia, and Europe, and are predominantly single-center and not publicly available. Such single-region, single-center datasets may embed systematic biases related to patient demographics, scanner types, acquisition protocols, and surgical practices, thereby reducing the applicability of trained models to other populations or institutions. Moving forward, increased cross-regional collaboration and harmonization techniques such as domain adaptation and federated learning will be important to mitigate these biases\u003csup\u003e89,141\u003c/sup\u003e. Data accessibility represents an additional challenge. With 65.5% of datasets being private and unavailable for reuse, reproducibility is often limited, and fair method comparison remains challenging. Establishing open, multi-center, multimodal epilepsy surgery benchmarks, alongside broader sharing of pretrained models, and code, will be critical for accelerating progress and improving community confidence\u003csup\u003e68,93,129\u003c/sup\u003e. Beyond scale and accessibility, label quality also presents inherent limitations\u003csup\u003e145\u003c/sup\u003e. Ground-truth labels such as SOZ/EZ localization, FCD lesion masks, or HFO annotations frequently contain noise due to subjective expert interpretation, heterogeneous clinical standards, or reliance on proxy markers such as resection zones or short-term postoperative outcomes. Improved multi-rater consensus labeling, uncertainty modeling, and the use of long-term outcomes will be needed to strengthen the reliability of training and evaluation. Finally, data modality usage remains highly skewed. Electrophysiology modalities account for 52.8% of all datasets, whereas other rich information sources, such as PET, rs-fMRI, diffusion imaging, and especially clinical text, are rarely exploited. With recent advances in large language models, leveraging clinical narratives, reports, and semiology descriptions may open new opportunities for improving prediction accuracy and capturing patient-specific clinical context. Expanding beyond electrophysiology to more diverse data modalities will be an important direction for future work\u003csup\u003e21,118\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eIn terms of modeling, the field remains dominated by supervised CNN-based approaches. Most studies directly adapt existing architectures from general computer vision or signal processing, rather than models specifically tailored to the characteristics of multimodal neuroimaging, electrophysiology, or surgical decision-making. Although these models often produce strong results on small, single-center datasets, their robustness in multi-center, heterogeneous clinical environments remains largely unknown\u003csup\u003e6,106\u003c/sup\u003e. Moreover, there remains a notable absence of systematic comparisons across model families, such as CNNs, transformers, GNNs, hybrid architectures, or generative models, making it difficult to determine which design choices consistently translate across settings. Notably, large language models (LLMs) have recently been introduced into this field, primarily for interpreting clinical text for SOZ/EZ localization\u003csup\u003e21,86\u003c/sup\u003e. However, their application remains limited to unimodal textual reasoning. Future work should explore the potential of multimodal LLMs as decision-support agents capable of integrating imaging, electrophysiology, and clinical narratives. LLMs may also support auxiliary tasks, including EEG/MRI annotation, study design, and automated report generation. A common limitation is the lack of external validation. With 77.2% of studies reporting no external validation, performance estimates are likely inflated due to single-center bias. Strengthening cross-center and cross-region generalization should therefore be a central methodological priority. Multi-center external validation, domain adaptation, and harmonization strategies will be essential to ensure that models perform reliably across diverse patient populations and clinical environments\u003csup\u003e76,111\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003eEvaluation and translation evidence remain key bottlenecks for deploying epilepsy surgery AI in health systems. Most studies rely solely on automated quantitative metrics such as accuracy, AUC, or Dice scores. While these measures are essential for algorithmic benchmarking, they provide limited insight into whether a model improves clinically meaningful outcomes in epilepsy surgery. Incorporating clinically relevant endpoints, such as changes in surgical planning, accuracy of resection margin prediction, seizure-freedom rates, reduction in evaluation time, or improvements in clinician confidence, would provide a more comprehensive assessment of real-world utility\u003csup\u003e22,84\u003c/sup\u003e. Workflow integration is also uncommon. 89.6% of included studies operate as offline research prototypes, with only a small number developed into decision-support tools or semi-integrated systems. This translation gap reflects a broader disconnect between algorithm development and real clinical workflows, driven by challenges such as limited access to surgical environments, high barriers to software integration, compatibility issues with existing medical devices and IT infrastructure, unclear regulatory requirements, and the lack of collaboration between researchers, clinical engineers, and industry partners\u003csup\u003e80,84,120\u003c/sup\u003e. Moving forward, evaluations should measure real clinical impact rather than just accuracy, including effects on surgical planning, workflow efficiency, and clinician confidence. Furthermore, AI systems must be usable in real workflows, support clinician-in-the-loop interaction, and integrate smoothly with existing systems\u003csup\u003e73\u003c/sup\u003e.\u003c/p\u003e\n\u003cp\u003ePublication volume in epilepsy surgery AI, predominantly deep learning-based, has increased since 2018 with notable accelerations in 2020 and 2022. This growth likely reflects the maturation of deep learning frameworks, increased availability of open-source tools and datasets, and broader adoption of computational methods within the neuroscience and medical imaging communities. Despite this momentum, the field still lacks standardized benchmarks or shared evaluation pipelines, which limits reproducibility and makes it difficult to compare methodological advances across studies. The geographic distribution of contributing research groups is similarly imbalanced. Asia and North America account for most publications, while contributions from Europe are moderate and those from other regions remain limited. This imbalance raises concerns regarding model fairness and global applicability, as datasets predominantly represent specific populations, acquisition protocols, and surgical practices\u003csup\u003e86\u003c/sup\u003e. Addressing these disparities will require coordinated multi-center consortia, greater international data sharing, and wider adoption of federated learning to support collaborative model development without requiring data exchange. These efforts are essential for building globally representative datasets and developing more generalizable and equitable AI-enabled decision-support systems for epilepsy surgery.\u003c/p\u003e\n\u003cp\u003eTaken together, our findings indicate that epilepsy surgery AI is expanding but remains fragmented across the care pathway. Evidence is constrained by small, geographically skewed, and often non-public datasets, alongside limited external validation and scarce multi-center evaluation. Most models perform well in controlled research settings but remain untested in the heterogeneous, multi-center environments where real clinical decisions are made. Evaluation practices emphasize technical accuracy rather than clinical impact, and workflow-integrated deployment is uncommon. Addressing these gaps will require a transition toward full-pathway AI systems, large-scale multimodal datasets, advanced and interpretable modeling strategies, rigorous generalization testing, and stronger collaboration across technical, clinical, and industry domains. Fig. 6 summarizes these limitations and outlines future directions for developing reliable, equitable, and clinically actionable decision-support systems for epilepsy surgery.\u003c/p\u003e\n\u003cp\u003eThis survey also has several limitations. First, we restricted our search to English language publications, which may have excluded relevant studies reported in other languages. Second, our review focused on peer-reviewed articles and did not incorporate recent preprints, potentially missing emerging research trends. In summary, this scoping review provides a comprehensive and structured synthesis of AI-enabled decision support in epilepsy surgery. We examined how models are distributed across clinical stages and tasks, characterized the underlying datasets, analyzed modeling approaches and validation strategies, and assessed evaluation practices and workflow integration. We further identified key methodological, data-related, and translational limitations and outlined future directions needed to advance this field. Our findings highlight implementation-relevant gaps, including uneven pathway coverage, small and geographically concentrated datasets, limited external validation, and scarce workflow-integrated evaluation. We hope this work serves as a foundational reference for clinicians, researchers, and industry partners seeking to develop reliable, generalizable, and clinically impactful decision-support systems for epilepsy surgery.\u0026nbsp;\u003c/p\u003e"},{"header":"Methods","content":"\u003ch2\u003eData Source and Article Selection\u003c/h2\u003e\n\u003cp\u003eThis scoping review was reported in accordance with the PRISMA-ScR guidelines, and the study selection process is summarized in the PRISMA flow diagram (Fig. 7). The protocol for this scoping review was not prospectively registered. A systematic search was conducted across seven electronic databases, including Scopus, Web of Science, Embase, PubMed, IEEE Xplore, ACM Digital Library, and CINAHL, to identify studies on AI-enabled decision support in epilepsy surgery published between January 2018 and May 2025. The search strategy combined terms related to deep learning, epilepsy, and surgery. Specifically, search terms related to deep learning were \u0026quot;deep learning\u0026quot;, \u0026quot;machine learning\u0026quot;, \u0026quot;neural network*\u0026quot;, \u0026quot;artificial intelligence\u0026quot;, \u0026quot;AI\u0026quot;, \u0026quot;LLM*\u0026quot;, \u0026quot;large language model*\u0026quot;, \u0026quot;foundation model*\u0026quot;, \u0026quot;vision language model*\u0026quot;, \u0026quot;vision-language model\u0026quot;; search terms related to epilepsy were \u0026quot;epilep*\u0026quot;, and \u0026quot;seizure*\u0026quot;; and search terms related to surgery involved \u0026quot;surgery\u0026quot;, \u0026quot;surgical\u0026quot;, \u0026quot;presurgical\u0026quot;, \u0026quot;preoperative\u0026quot;, \u0026quot;intraoperative\u0026quot;, \u0026quot;postoperative\u0026quot;. Detailed search queries and Boolean operators (AND/OR) for each database are provided in the Supplementary Table 1. Study selection followed a two-stage screening process. First, two reviewers independently screened titles and abstracts using predefined inclusion and exclusion criteria. Second, full texts of potentially eligible studies were independently evaluated, with disagreements resolved through discussion or, if necessary, adjudication by a third reviewer. Specifically, studies were included if they met the following criteria: (1) empirical investigations of deep learning-based methods; (2) applications directly related to epilepsy surgery, including pre-operative, intra-operative, or post-operative stages. Conversely, studies were excluded if they: (1) relied solely on traditional machine learning methods; (2) focused on epilepsy-related applications that were not connected to the surgical pathway; (3) solely for surgical candidate screening or pre-surgical triage without involving tasks directly related to the operative workflow, or (4) were non-English publications, review articles, editorials, or abstracts without accessible full text.\u0026nbsp;\u003c/p\u003e\n\u003ch2\u003eData Extraction and Analysis\u003c/h2\u003e\n\u003cp\u003eData from the included studies were systematically extracted using a predefined template to capture key information (Supplementary Table 2). Extracted items included: (1) basic study information, including title, publication type, published year, and author region (based on the first author\u0026rsquo;s primary institutional affiliation: Asia, Europe, North America, Oceania, or Africa); (2) epilepsy surgery stage \u0026amp; clinical task, specifying the targeted surgical stage (pre-operative, intra-operative, or post-operative), and the corresponding clinical task; (3) data characteristics, including data source region, data modality, data size, and dataset accessibility (public, private, or mixed). In this survey, data size was recorded as the number of patients whose data were used in the study, including both individuals with epilepsy and healthy controls; (4) modeling details, including deep model architecture and training strategy (e.g., supervised learning, transfer learning, self-supervised learning, weak supervision, federated learning, or other specified strategies); (5) evaluation, including external validation, evaluation type (manual, automatic, or mixed), reported evaluation metrics, and the primary performance summary; and (6) clinical translation, in which we categorized the deployment level of each deep learning system as offline research (models evaluated retrospectively without clinical use), standalone decision support tools (outputs available for clinician review but not embedded), or semi-integrated clinical systems (outputs partially incorporated into clinical workflows). All data extraction was performed independently by two reviewers, with discrepancies resolved through discussion and consensus.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eThe basic information, clinical context, dataset characteristics, methodological details, and translational relevance of the included deep learning studies were systematically documented and categorized using our taxonomy. We synthesized these data to highlight key patterns in AI-enabled decision support systems for epilepsy surgery. We first summarized the survey scope, including surgery stage, clinical tasks, data modalities, and model architectures. We then analyzed study metadata such as geographic distribution, data accessibility, and training strategies. Finally, we reviewed evaluation practices and clinical translation stages and outlined major limitations and future directions.\u003c/p\u003e"},{"header":"Declarations","content":"\u003ch2\u003eData Availability\u003c/h2\u003e\n\u003cp\u003eAll data generated during this study are presented in the Supplementary Materials.\u003c/p\u003e\n\u003ch2\u003eCode Availability\u003c/h2\u003e\n\u003cp\u003eNo code was generated during this study.\u003c/p\u003e\n\u003ch2\u003eAcknowledgments\u003c/h2\u003e\n\u003cp\u003eThis work was supported by the National Institutes of Health\u0026rsquo;s National Center for Complementary and Integrative Health under grant number R01AT009457, National Institute on Aging under grant number R01AG078154, and National Cancer Institute under grant number R01CA287413. The content is solely the responsibility of the authors and does not represent the official views of the National Institutes of Health.\u0026nbsp;We thank Janie Sha and Yitao Cheng (Wayzata High School) for assistance with literature searching and reference organization.\u003c/p\u003e\n\u003ch2\u003eAuthor Contributions\u003c/h2\u003e\n\u003cp\u003eK.Y. and R.Z. conceptualized the study and led the work. K.Y. and S.Z. searched papers. K.Y., S.Z., M.S., Z.Z., Y.H, Y.S., M.Z., and B.Y. conducted paper screening and data extraction. K.Y. performed data synthesis and contributed to the writing. K.Y., F.L., S. P., Z. S., M.L., and R.Z. revised the manuscript. R.Z. supervised the study. All authors read and approved the final version.\u003c/p\u003e\n\u003ch2\u003eCompeting Interests\u003c/h2\u003e\n\u003cp\u003eThe authors declare no competing interests.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eWorld Health Organization. Epilepsy: Key facts. \u003cem\u003eWorld Health Organization\u003c/em\u003e https://www.who.int/news-room/fact-sheets/detail/epilepsy (2024).\u003c/li\u003e\n\u003cli\u003eVaughan, K. A. \u003cem\u003eet al.\u003c/em\u003e An estimation of global volume of surgically treatable epilepsy based on a systematic review and meta-analysis of epilepsy. \u003cem\u003eJ. Neurosurg.\u003c/em\u003e \u003cstrong\u003e130\u003c/strong\u003e, 1127\u0026ndash;1141 (2019).\u003c/li\u003e\n\u003cli\u003eThijs, R. D., Surges, R., O\u0026rsquo;Brien, T. J. \u0026amp; Sander, J. W. Epilepsy in adults. \u003cem\u003eLancet\u003c/em\u003e \u003cstrong\u003e393\u003c/strong\u003e, 689\u0026ndash;701 (2019).\u003c/li\u003e\n\u003cli\u003eRyvlin, P., Cross, J. H. \u0026amp; Rheims, S. Epilepsy surgery in children and adults. \u003cem\u003eLancet Neurol.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 1114\u0026ndash;1126 (2014).\u003c/li\u003e\n\u003cli\u003eRosenow, F. \u0026amp; L\u0026uuml;ders, H. Presurgical evaluation of epilepsy. \u003cem\u003eBrain : a journal of neurology\u003c/em\u003e \u003cstrong\u003e124 Pt 9\u003c/strong\u003e, 1683\u0026ndash;1700 (2001).\u003c/li\u003e\n\u003cli\u003eIto, Y. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based diagnosis of temporal lobe epilepsy associated with hippocampal sclerosis: An MRI study. \u003cem\u003eEpilepsy Res.\u003c/em\u003e \u003cstrong\u003e178\u003c/strong\u003e, 106815 (2021).\u003c/li\u003e\n\u003cli\u003eGill, R. S. \u003cem\u003eet al.\u003c/em\u003e Multicenter validation of a deep learning detection algorithm for focal cortical dysplasia. \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e97\u003c/strong\u003e, e1571\u0026ndash;e1582 (2021).\u003c/li\u003e\n\u003cli\u003eNiyas, S. \u003cem\u003eet al.\u003c/em\u003e Segmentation of focal cortical dysplasia lesions from magnetic resonance images using 3D convolutional neural networks. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e \u003cstrong\u003e70\u003c/strong\u003e, 102951 (2021).\u003c/li\u003e\n\u003cli\u003eDaoud, H. \u0026amp; Bayoumi, M. Deep learning approach for epileptic focus localization. \u003cem\u003eIEEE Trans. Biomed. Circuits Syst.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 209\u0026ndash;220 (2020).\u003c/li\u003e\n\u003cli\u003eXiao, L. \u003cem\u003eet al.\u003c/em\u003e Automatic localization of seizure onset zone from high-frequency SEEG signals: A preliminary study. \u003cem\u003eIEEE J. Transl. Eng. Health Med.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 1\u0026ndash;10 (2021).\u003c/li\u003e\n\u003cli\u003eSaminu, S. \u003cem\u003eet al.\u003c/em\u003e Application of deep learning and WT-SST in localization of epileptogenic zone using epileptic EEG signals. \u003cem\u003eAppl. Sci. (Basel)\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 4879 (2022).\u003c/li\u003e\n\u003cli\u003eAhmedt-Aristizabal, D. \u003cem\u003eet al.\u003c/em\u003e Deep facial analysis: A new phase I epilepsy evaluation using computer vision. \u003cem\u003eEpilepsy Behav.\u003c/em\u003e \u003cstrong\u003e82\u003c/strong\u003e, 17\u0026ndash;24 (2018).\u003c/li\u003e\n\u003cli\u003eLee, M.-H. \u003cem\u003eet al.\u003c/em\u003e Novel deep learning network analysis of electrical stimulation mapping-driven diffusion MRI tractography to improve preoperative evaluation of pediatric epilepsy. \u003cem\u003eIEEE Trans. Biomed. Eng.\u003c/em\u003e \u003cstrong\u003e67\u003c/strong\u003e, 3151\u0026ndash;3162 (2020).\u003c/li\u003e\n\u003cli\u003eGranados, A. \u003cem\u003eet al.\u003c/em\u003e Towards uncertainty quantification for electrode bending prediction in stereotactic neurosurgery. in \u003cem\u003e2020 IEEE 17th International Symposium on Biomedical Imaging (ISBI)\u003c/em\u003e 674\u0026ndash;677 (IEEE, 2020).\u003c/li\u003e\n\u003cli\u003eLiu, H., Holloway, K. L., Englot, D. J. \u0026amp; Dawant, B. M. A multi-rater comparative study of automatic target localization methods for epilepsy deep Brain Stimulation procedures. \u003cem\u003earXiv [eess.IV]\u003c/em\u003e (2022) doi:10.1117/12.2612927.\u003c/li\u003e\n\u003cli\u003eNejedly, P. \u003cem\u003eet al.\u003c/em\u003e Leveraging interictal multimodal features and graph neural networks for automated planning of epilepsy surgery. \u003cem\u003eBrain Commun.\u003c/em\u003e \u003cstrong\u003e7\u003c/strong\u003e, fcaf140 (2025).\u003c/li\u003e\n\u003cli\u003eConstantino, A. C. \u003cem\u003eet al.\u003c/em\u003e Expert-level intracranial electroencephalogram ictal pattern detection by a deep learning neural network. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 603868 (2021).\u003c/li\u003e\n\u003cli\u003eTang, Y. \u003cem\u003eet al.\u003c/em\u003e Machine learning-derived multimodal neuroimaging of presurgical target area to predict individual\u0026rsquo;s seizure outcomes after epilepsy surgery. \u003cem\u003eFront. Cell Dev. Biol.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 669795 (2021).\u003c/li\u003e\n\u003cli\u003eSun, L. \u003cem\u003eet al.\u003c/em\u003e High-performance prediction of epilepsy surgical outcomes based on the genetic neural networks and hybrid iEEG marker. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 6198 (2024).\u003c/li\u003e\n\u003cli\u003eLee, M.-H. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based tract classification of preoperative DWI tractography advances the prediction of short-term postoperative language improvement in children with drug-resistant epilepsy. \u003cem\u003eIEEE Trans. Biomed. Eng.\u003c/em\u003e \u003cstrong\u003e72\u003c/strong\u003e, 565\u0026ndash;576 (2025).\u003c/li\u003e\n\u003cli\u003eChiang, K.-L. \u003cem\u003eet al.\u003c/em\u003e Customized GPT model largely increases surgery decision accuracy for pharmaco-resistant epilepsy. \u003cem\u003eJ. Clin. Neurosci.\u003c/em\u003e \u003cstrong\u003e130\u003c/strong\u003e, 110918 (2024).\u003c/li\u003e\n\u003cli\u003eArnold, T. C. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based automated segmentation of resection cavities on postsurgical epilepsy MRI. \u003cem\u003eNeuroImage Clin.\u003c/em\u003e \u003cstrong\u003e36\u003c/strong\u003e, 103154 (2022).\u003c/li\u003e\n\u003cli\u003eSun, R., Sohrabpour, A., Worrell, G. A. \u0026amp; He, B. Deep neural networks constrained by neural mass models improve electrophysiological source imaging of spatiotemporal brain dynamics. \u003cem\u003eProc. Natl. Acad. Sci. U. S. A.\u003c/em\u003e \u003cstrong\u003e119\u003c/strong\u003e, e2201128119 (2022).\u003c/li\u003e\n\u003cli\u003eChanra, V. \u003cem\u003eet al.\u003c/em\u003e Development and prospective clinical validation of a convolutional neural network for automated detection and segmentation of focal cortical dysplasias. \u003cem\u003eEpilepsy Res.\u003c/em\u003e \u003cstrong\u003e202\u003c/strong\u003e, 107357 (2024).\u003c/li\u003e\n\u003cli\u003eZhang, X., Zhang, X., Huang, Q. \u0026amp; Chen, F. A review of epilepsy detection and prediction methods based on EEG signal processing and deep learning. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eSi, Y. Machine learning applications for electroencephalograph signals in epilepsy: a quick review. \u003cem\u003eActa Epileptologica\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, (2020).\u003c/li\u003e\n\u003cli\u003eNafea, M. S. \u0026amp; Ismail, Z. H. Supervised Machine learning and Deep learning techniques for epileptic seizure recognition using EEG signals-A systematic literature review. \u003cem\u003eBioengineering (Basel)\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 781 (2022).\u003c/li\u003e\n\u003cli\u003eRasheed, K. \u003cem\u003eet al.\u003c/em\u003e Machine Learning for Predicting Epileptic Seizures Using EEG Signals: A Review. \u003cem\u003eIEEE Rev. Biomed. Eng.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 139\u0026ndash;155 (2021).\u003c/li\u003e\n\u003cli\u003eQamar, W. U. R., Lee, M.-H. \u0026amp; Abibullaev, B. Deep learning in intracranial EEG for seizure detection: advances, challenges, and clinical applications. \u003cem\u003eFrontiers in Neuroscience\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 1677898 (2025).\u003c/li\u003e\n\u003cli\u003eDaida, A. \u003cem\u003eet al.\u003c/em\u003e AI ‐based localization of the epileptogenic zone using intracranial EEG. \u003cem\u003eEpilepsia Open\u003c/em\u003e (2025) doi:10.1002/epi4.70130.\u003c/li\u003e\n\u003cli\u003eBerger, M., Licandro, R., Nenning, K.-H., Langs, G. \u0026amp; Bonelli, S. B. Artificial intelligence applied to epilepsy imaging: Current status and future perspectives. \u003cem\u003eRevue Neurologique\u003c/em\u003e \u003cstrong\u003e181\u003c/strong\u003e, 420\u0026ndash;424 (2025).\u003c/li\u003e\n\u003cli\u003eSone, D. \u0026amp; Beheshti, I. Clinical application of machine learning models for brain imaging in epilepsy: A review. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 684825 (2021).\u003c/li\u003e\n\u003cli\u003eYuan, J. \u003cem\u003eet al.\u003c/em\u003e Machine learning applications on neuroimaging for diagnosis and prognosis of epilepsy: A review. \u003cem\u003eJournal of Neuroscience Methods\u003c/em\u003e \u003cstrong\u003e368\u003c/strong\u003e, 109441 (2022).\u003c/li\u003e\n\u003cli\u003eTan, S. \u003cem\u003eet al.\u003c/em\u003e Epilepsy surgery candidate identification with artificial intelligence: An implementation study. \u003cem\u003eJ. Clin. Neurosci.\u003c/em\u003e \u003cstrong\u003e135\u003c/strong\u003e, 111144 (2025).\u003c/li\u003e\n\u003cli\u003eChowdhury, S. \u003cem\u003eet al.\u003c/em\u003e Using large language models to identify candidates for pediatric epilepsy surgery (P3-8.001). \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e102\u003c/strong\u003e, (2024).\u003c/li\u003e\n\u003cli\u003eHao, Y., Khoo, H. M., von Ellenrieder, N., Zazubovits, N. \u0026amp; Gotman, J. DeepIED: An epileptic discharge detector for EEG-fMRI based on deep learning. \u003cem\u003eNeuroImage Clin.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 962\u0026ndash;975 (2018).\u003c/li\u003e\n\u003cli\u003eRaghu, S., Sriraam, N., Temel, Y., Rao, S. V. \u0026amp; Kubben, P. L. EEG based multi-class seizure type classification using convolutional neural network and transfer learning. \u003cem\u003eNeural Netw.\u003c/em\u003e \u003cstrong\u003e124\u003c/strong\u003e, 202\u0026ndash;212 (2020).\u003c/li\u003e\n\u003cli\u003eAbou Jaoude, M. \u003cem\u003eet al.\u003c/em\u003e Noninvasive detection of hippocampal epileptiform activity on scalp electroencephalogram. \u003cem\u003eJAMA Neurol.\u003c/em\u003e \u003cstrong\u003e79\u003c/strong\u003e, 614\u0026ndash;622 (2022).\u003c/li\u003e\n\u003cli\u003eYu, Z. \u003cem\u003eet al.\u003c/em\u003e Electrophysiological brain imaging based on simulation-driven deep learning in the context of epilepsy. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e285\u003c/strong\u003e, 120490 (2024).\u003c/li\u003e\n\u003cli\u003eVajiram, J., Sivakumar, Jena, R. \u0026amp; Maurya, U. Epilepsy detection by different modalities with the use of AI-assisted models. \u003cem\u003eArtificial Intelligence and Applications\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 233\u0026ndash;246 (2023).\u003c/li\u003e\n\u003cli\u003eKuang, Z. \u003cem\u003eet al.\u003c/em\u003e Seizure onset zone detection based on convolutional neural networks and EEG signals. \u003cem\u003eBrain Sci.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1090 (2024).\u003c/li\u003e\n\u003cli\u003eDokare, I. \u0026amp; Gupta, S. Brain-region specific epileptic seizure detection through EEG dynamics: integrating spectral features, SMOTE and long short-term memory networks. \u003cem\u003eCogn. Neurodyn.\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 67 (2025).\u003c/li\u003e\n\u003cli\u003eLiu, Y., Wang, Y. \u0026amp; Wang, T. Non-invasive localization of epileptogenic zone in drug-resistant epilepsy based on time-frequency analysis and VGG convolutional neural network. \u003cem\u003eBioengineering (Basel)\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 443 (2025).\u003c/li\u003e\n\u003cli\u003eMedvedev, A., Agoureeva, G. I. \u0026amp; Murro, A. A Long Short-Term Memory neural network for the detection of epileptiform spikes and high frequency oscillations. \u003cem\u003eSci. Rep.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, (2019).\u003c/li\u003e\n\u003cli\u003eMa, K. \u003cem\u003eet al.\u003c/em\u003e Automatic detection of high frequency oscillations (80-500Hz) based on convolutional neural network in human intracerebral electroencephalogram. \u003cem\u003eAnnu. Int. Conf. IEEE Eng. Med. Biol. Soc.\u003c/em\u003e \u003cstrong\u003e2019\u003c/strong\u003e, 5133\u0026ndash;5136 (2019).\u003c/li\u003e\n\u003cli\u003eSui, L., Zhao, X., Zhao, Q., Tanaka, T. \u0026amp; Cao, J. Localization of epileptic foci by using convolutional neural network based on iEEG. in \u003cem\u003eIFIP Advances in Information and Communication Technology\u003c/em\u003e 331\u0026ndash;339 (Springer International Publishing, Cham, 2019).\u003c/li\u003e\n\u003cli\u003eBajaj, V., Taran, S., Tanyildizi, E. \u0026amp; Sengur, A. Robust approach based on convolutional neural networks for identification of focal EEG signals. \u003cem\u003eIEEE Sens. Lett.\u003c/em\u003e \u003cstrong\u003e3\u003c/strong\u003e, 1\u0026ndash;4 (2019).\u003c/li\u003e\n\u003cli\u003eDaoud, H. \u0026amp; Bayoumi, M. Semi-supervised learning for epileptic focus localization using deep convolutional autoencoder. in \u003cem\u003e2019 IEEE Biomedical Circuits and Systems Conference (BioCAS)\u003c/em\u003e 1\u0026ndash;4 (IEEE, 2019).\u003c/li\u003e\n\u003cli\u003eMadhavan, S., Tripathy, R. K. \u0026amp; Pachori, R. B. Time-frequency domain deep convolutional neural network for the classification of focal and non-focal EEG signals. \u003cem\u003eIEEE Sens. J.\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, 3078\u0026ndash;3086 (2020).\u003c/li\u003e\n\u003cli\u003eDaoud, H. \u0026amp; Bayoumi, M. A convolutional gated recurrent neural network for seizure onset localization. in \u003cem\u003e2020 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)\u003c/em\u003e 2572\u0026ndash;2576 (IEEE, 2020).\u003c/li\u003e\n\u003cli\u003eZhao, X. \u003cem\u003eet al.\u003c/em\u003e Classification of epileptic IEEG signals by CNN and data augmentation. in \u003cem\u003eICASSP 2020 - 2020 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)\u003c/em\u003e 926\u0026ndash;930 (IEEE, 2020).\u003c/li\u003e\n\u003cli\u003eFraiwan, L. \u0026amp; Alkhodari, M. Classification of focal and non-focal epileptic patients using single channel EEG and long short-term memory learning system. \u003cem\u003eIEEE Access\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 77255\u0026ndash;77262 (2020).\u003c/li\u003e\n\u003cli\u003eWang, Z. \u0026amp; Li, C. Classifying cross-frequency coupling pattern in epileptogenic tissues by convolutional neural network. \u003cem\u003eAnnu. Int. Conf. IEEE Eng. Med. Biol. Soc.\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, 3440\u0026ndash;3443 (2020).\u003c/li\u003e\n\u003cli\u003eChung, Y. G. \u003cem\u003eet al.\u003c/em\u003e Deep convolutional neural network based interictal-preictal electroencephalography prediction: Application to focal cortical dysplasia type-II. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 594679 (2020).\u003c/li\u003e\n\u003cli\u003eZhao, X., Sui, L., Tanaka, T. \u0026amp; Cao, J. Epileptic Focus Localization Based on iEEG Plot Images by Using Convolutional Neural Network.\u003c/li\u003e\n\u003cli\u003eQin, H., Wu, M., Wan, X. \u0026amp; Du, Y. Stacked sparse autoencoder based automatic detection of ripples and fast ripples in epilepsy. in \u003cem\u003e2020 39th Chinese Control Conference (CCC)\u003c/em\u003e 2833\u0026ndash;2837 (IEEE, 2020).\u003c/li\u003e\n\u003cli\u003eWang, Y. \u003cem\u003eet al.\u003c/em\u003e Computer-aided intracranial EEG signal identification method based on a multi-branch deep learning fusion model and clinical validation. \u003cem\u003eBrain Sci.\u003c/em\u003e \u003cstrong\u003e11\u003c/strong\u003e, 615 (2021).\u003c/li\u003e\n\u003cli\u003eXia, M., Sui, L., Zhao, X., Tanaka, T. \u0026amp; Cao, J. Convolution neural network recognition of epileptic foci based on composite signal processing of electroencephalograph data. \u003cem\u003eProcedia Comput. Sci.\u003c/em\u003e \u003cstrong\u003e192\u003c/strong\u003e, 688\u0026ndash;696 (2021).\u003c/li\u003e\n\u003cli\u003eYamamoto, S. \u003cem\u003eet al.\u003c/em\u003e Data-driven electrophysiological feature based on deep learning to detect epileptic seizures. \u003cem\u003eJ. Neural Eng.\u003c/em\u003e \u003cstrong\u003e18\u003c/strong\u003e, 056040 (2021).\u003c/li\u003e\n\u003cli\u003eTakayanagi, Y., Takayama, Y., Iijima, K., Iwasaki, M. \u0026amp; Ono, Y. Efficient detection of high-frequency biomarker signals of epilepsy by a transfer-learning-based convolutional neural network. \u003cem\u003eAdv. Biomed. Eng.\u003c/em\u003e \u003cstrong\u003e10\u003c/strong\u003e, 158\u0026ndash;165 (2021).\u003c/li\u003e\n\u003cli\u003eModak, S., Singha Roy, S., Bose, R. \u0026amp; Chatterjee, S. Focal epileptic area recognition employing cross EEG rhythm spectrum images and convolutional neural network. \u003cem\u003eIEEE Sens. J.\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 23335\u0026ndash;23343 (2021).\u003c/li\u003e\n\u003cli\u003eDaoud, H. \u0026amp; Bayoumi, M. Generative adversarial network based semi-supervised learning for epileptic focus localization. in \u003cem\u003e2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)\u003c/em\u003e 1281\u0026ndash;1285 (IEEE, 2021).\u003c/li\u003e\n\u003cli\u003eSui, L., Zhao, X., Zhao, Q., Tanaka, T. \u0026amp; Cao, J. Hybrid convolutional neural network for localization of epileptic focus based on iEEG. \u003cem\u003eNeural Plast.\u003c/em\u003e \u003cstrong\u003e2021\u003c/strong\u003e, 6644365 (2021).\u003c/li\u003e\n\u003cli\u003eRen, G. \u003cem\u003eet al.\u003c/em\u003e Identification of epileptogenic and non-epileptogenic high-frequency oscillations using a multi-feature convolutional neural network model. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 640526 (2021).\u003c/li\u003e\n\u003cli\u003eGuo, J., Wang, Y., Yang, Y. \u0026amp; Kang, G. IEEG-TCN: A concise and robust temporal convolutional network for intracranial electroencephalogram signal identification. in \u003cem\u003e2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM)\u003c/em\u003e 668\u0026ndash;673 (IEEE, 2021).\u003c/li\u003e\n\u003cli\u003eZhao, X., Sol\u0026eacute;-Casals, J., Zhao, Q., Cao, J. \u0026amp; Tanaka, T. Multi-feature fusion for epileptic focus localization based on tensor representation. in \u003cem\u003e2021 Asia-Pacific Signal and Information Processing Association Annual Summit and Conference (APSIPA ASC)\u003c/em\u003e 1323\u0026ndash;1327 (IEEE, 2021).\u003c/li\u003e\n\u003cli\u003eWang, Y. \u003cem\u003eet al.\u003c/em\u003e A cross-patient SEEG epileptic signal detection method based on adaptive feature fusion of brain network features and single-channel features. in \u003cem\u003eProceedings of the 2022 9th International Conference on Biomedical and Bioinformatics Engineering\u003c/em\u003e 165\u0026ndash;172 (ACM, New York, NY, USA, 2022).\u003c/li\u003e\n\u003cli\u003eWang, Y. \u003cem\u003eet al.\u003c/em\u003e Automatic localization of seizure onset zone based on multi-epileptogenic biomarkers analysis of single-contact from interictal SEEG. \u003cem\u003eBioengineering (Basel)\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, (2022).\u003c/li\u003e\n\u003cli\u003eChen, J. \u003cem\u003eet al.\u003c/em\u003e BrainNet: Epileptic wave detection from SEEG with hierarchical graph diffusion learning. in \u003cem\u003eProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining\u003c/em\u003e 2741\u0026ndash;2751 (ACM, New York, NY, USA, 2022).\u003c/li\u003e\n\u003cli\u003eZhao, X. \u003cem\u003eet al.\u003c/em\u003e A multi-head self-attention deep learning approach for detection and recommendation of neuromagnetic high frequency oscillations in epilepsy. \u003cem\u003eFront. Neuroinform.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 771965 (2022).\u003c/li\u003e\n\u003cli\u003eZhang, M., Liu, J., Liu, C., Wu, T. \u0026amp; Peng, X. An efficient CADNet for classification of high-frequency oscillations in magnetoencephalography. in \u003cem\u003e2022 4th International Conference on Robotics and Computer Vision (ICRCV)\u003c/em\u003e (IEEE, 2022). doi:10.1109/icrcv55858.2022.9953255.\u003c/li\u003e\n\u003cli\u003eGuo, J. \u003cem\u003eet al.\u003c/em\u003e Transformer-based high-frequency oscillation signal detection on magnetoencephalography from epileptic patients. \u003cem\u003eFront. Mol. Biosci.\u003c/em\u003e \u003cstrong\u003e9\u003c/strong\u003e, 822810 (2022).\u003c/li\u003e\n\u003cli\u003eZheng, L. \u003cem\u003eet al.\u003c/em\u003e An artificial intelligence-based pipeline for automated detection and localisation of epileptic sources from magnetoencephalography. \u003cem\u003eJ. Neural Eng.\u003c/em\u003e \u003cstrong\u003e20\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eSun, R., Zhang, W., Bagić, A. \u0026amp; He, B. Deep learning based source imaging provides strong sublobar localization of epileptogenic zone from MEG interictal spikes. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e281\u003c/strong\u003e, 120366 (2023).\u003c/li\u003e\n\u003cli\u003eGuo, J. \u003cem\u003eet al.\u003c/em\u003e A stacked sparse autoencoder-based detector for automatic identification of neuromagnetic high frequency oscillations in epilepsy. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e \u003cstrong\u003e37\u003c/strong\u003e, 2474\u0026ndash;2482 (2018).\u003c/li\u003e\n\u003cli\u003eLiu, J. \u003cem\u003eet al.\u003c/em\u003e A novel MEGNet for classification of high-frequency oscillations in magnetoencephalography of epileptic patients. \u003cem\u003eComplexity\u003c/em\u003e \u003cstrong\u003e2020\u003c/strong\u003e, 1\u0026ndash;9 (2020).\u003c/li\u003e\n\u003cli\u003eGuo, J. \u003cem\u003eet al.\u003c/em\u003e Automatic and accurate epilepsy ripple and fast ripple detection via virtual sample generation and attention neural networks. \u003cem\u003eIEEE Trans. Neural Syst. Rehabil. Eng.\u003c/em\u003e \u003cstrong\u003e28\u003c/strong\u003e, 1710\u0026ndash;1719 (2020).\u003c/li\u003e\n\u003cli\u003ePark, K.-I. \u003cem\u003eet al.\u003c/em\u003e Lateralizing value of artificial intelligence-based segmentation software in MRI-negative focal epilepsy. \u003cem\u003eJ. Epilepsy Res.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 59\u0026ndash;65 (2024).\u003c/li\u003e\n\u003cli\u003eKaestner, E. \u003cem\u003eet al.\u003c/em\u003e Convolutional neural network algorithm to determine lateralization of seizure onset in patients with epilepsy: A proof-of-principle study: A proof-of-principle study. \u003cem\u003eNeurology\u003c/em\u003e \u003cstrong\u003e101\u003c/strong\u003e, e324\u0026ndash;e335 (2023).\u003c/li\u003e\n\u003cli\u003eFlaus, A. \u003cem\u003eet al.\u003c/em\u003e Deep-learning predicted PET can be subtracted from the true clinical fluorodeoxyglucose PET co-registered to MRI to identify the epileptogenic zone in focal epilepsy. \u003cem\u003eEpilepsia Open\u003c/em\u003e \u003cstrong\u003e8\u003c/strong\u003e, 1440\u0026ndash;1451 (2023).\u003c/li\u003e\n\u003cli\u003eLuckett, P. H. \u003cem\u003eet al.\u003c/em\u003e Deep learning resting state functional magnetic resonance imaging lateralization of temporal lobe epilepsy. \u003cem\u003eEpilepsia\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e, 1542\u0026ndash;1552 (2022).\u003c/li\u003e\n\u003cli\u003eNandakumar, N., Hsu, D., Ahmed, R. \u0026amp; Venkataraman, A. DeepEZ: A graph convolutional network for automated epileptogenic zone localization from resting-state fMRI connectivity. \u003cem\u003eIEEE Trans. Biomed. Eng.\u003c/em\u003e \u003cstrong\u003e70\u003c/strong\u003e, 216\u0026ndash;227 (2023).\u003c/li\u003e\n\u003cli\u003eNandakumar, N., Hsu, D., Ahmed, R. \u0026amp; Venkataraman, A. A deep learning framework to characterize noisy labels in epileptogenic zone localization using functional connectivity. \u003cem\u003eProc. IEEE Int. Symp. Biomed. Imaging\u003c/em\u003e \u003cstrong\u003e2024\u003c/strong\u003e, 1\u0026ndash;5 (2024).\u003c/li\u003e\n\u003cli\u003eKamboj, P., Banerjee, A. \u0026amp; Gupta, S. K. S. Expert knowledge driven human-AI collaboration for medical imaging: A study on epileptic seizure onset zone identification. \u003cem\u003eIEEE Trans. Artif. Intell.\u003c/em\u003e \u003cstrong\u003e5\u003c/strong\u003e, 5352\u0026ndash;5368 (2024).\u003c/li\u003e\n\u003cli\u003eKamboj, P., Banerjee, A., Boerwinkle, V. L. \u0026amp; Gupta, S. K. S. The expert\u0026rsquo;s knowledge combined with AI outperforms AI alone in seizure onset zone localization using resting state fMRI. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 1324461 (2023).\u003c/li\u003e\n\u003cli\u003eLuo, Y. \u003cem\u003eet al.\u003c/em\u003e Clinical value of ChatGPT for epilepsy presurgical decision-making: Systematic evaluation of seizure semiology interpretation. \u003cem\u003eJ. Med. Internet Res.\u003c/em\u003e \u003cstrong\u003e27\u003c/strong\u003e, e69173 (2025).\u003c/li\u003e\n\u003cli\u003eLi, Z. \u003cem\u003eet al.\u003c/em\u003e Deep learning of simultaneous intracranial and scalp EEG for prediction, detection, and lateralization of mesial temporal lobe seizures. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 705119 (2021).\u003c/li\u003e\n\u003cli\u003eYang, Y., Li, F., Luo, J., Qin, X. \u0026amp; Huang, D. Epileptic focus localization using transfer learning on multi-modal EEG. \u003cem\u003eFront. Comput. Neurosci.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 1294770 (2023).\u003c/li\u003e\n\u003cli\u003eJeong, J.-W. \u003cem\u003eet al.\u003c/em\u003e Multi-scale deep learning of clinically acquired multi-modal MRI improves the localization of seizure onset zone in children with drug-resistant epilepsy. \u003cem\u003eIEEE J. Biomed. Health Inform.\u003c/em\u003e \u003cstrong\u003e26\u003c/strong\u003e, 5529\u0026ndash;5539 (2022).\u003c/li\u003e\n\u003cli\u003eFard, A. S. \u003cem\u003eet al.\u003c/em\u003e Image synthesis of interictal SPECT from MRI and PET using machine learning. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1383773 (2024).\u003c/li\u003e\n\u003cli\u003eBanerjee, S. \u003cem\u003eet al.\u003c/em\u003e Localization of epileptogenicity using incomplete MRI sequence data in children with seizure onset zone involving temporal lobe. in \u003cem\u003e2024 IEEE International Symposium on Biomedical Imaging (ISBI)\u003c/em\u003e vol. 2 1\u0026ndash;5 (IEEE, 2024).\u003c/li\u003e\n\u003cli\u003eZotova, D., Jung, J. \u0026amp; Lartizien, C. GAN-based synthetic FDG PET images from T1 brain MRI can serve to improve performance of deep unsupervised anomaly detection models. in \u003cem\u003eSimulation and Synthesis in Medical Imaging\u003c/em\u003e 142\u0026ndash;152 (Springer International Publishing, Cham, 2021).\u003c/li\u003e\n\u003cli\u003eJiang, J. \u003cem\u003eet al.\u003c/em\u003e Automated detection of hippocampal sclerosis using real-world clinical MRI images. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 1180679 (2023).\u003c/li\u003e\n\u003cli\u003eWang, H., Ahmed, S. N. \u0026amp; Mandal, M. Automated detection of focal cortical dysplasia using a deep convolutional neural network. \u003cem\u003eComput. Med. Imaging Graph.\u003c/em\u003e \u003cstrong\u003e79\u003c/strong\u003e, 101662 (2020).\u003c/li\u003e\n\u003cli\u003eZhang, D. \u003cem\u003eet al.\u003c/em\u003e Deep learning and radiomics based automatic diagnosis of hippocampal sclerosis. \u003cem\u003eInt. J. Neurosci.\u003c/em\u003e \u003cstrong\u003e133\u003c/strong\u003e, 947\u0026ndash;958 (2023).\u003c/li\u003e\n\u003cli\u003eKaestner, E. \u003cem\u003eet al.\u003c/em\u003e Adding the third dimension: 3D convolutional neural network diagnosis of temporal lobe epilepsy. \u003cem\u003eBrain Commun.\u003c/em\u003e \u003cstrong\u003e6\u003c/strong\u003e, fcae346 (2024).\u003c/li\u003e\n\u003cli\u003eZhang, S. \u003cem\u003eet al.\u003c/em\u003e Deep learning-based automated lesion segmentation on pediatric focal cortical dysplasia II preoperative MRI: a reliable approach. \u003cem\u003eInsights Imaging\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 71 (2024).\u003c/li\u003e\n\u003cli\u003eBijay Dev, K. M. \u003cem\u003eet al.\u003c/em\u003e Automatic detection and localization of Focal Cortical Dysplasia lesions in MRI using fully convolutional neural network. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e \u003cstrong\u003e52\u003c/strong\u003e, 218\u0026ndash;225 (2019).\u003c/li\u003e\n\u003cli\u003eFeng, C., Zhao, H., Li, Y. \u0026amp; Wen, J. Automatic localization and segmentation of focal cortical dysplasia in FLAIR-negative patients using a convolutional neural network. \u003cem\u003eJ. Appl. Clin. Med. Phys.\u003c/em\u003e \u003cstrong\u003e21\u003c/strong\u003e, 215\u0026ndash;226 (2020).\u003c/li\u003e\n\u003cli\u003eAminpour, A., Ebrahimi, M. \u0026amp; Widjaja, E. Lesion localization in paediatric epilepsy using patch-based convolutional neural network. in \u003cem\u003eLecture Notes in Computer Science\u003c/em\u003e 216\u0026ndash;227 (Springer International Publishing, Cham, 2020).\u003c/li\u003e\n\u003cli\u003eAlaverdyan, Z., Jung, J., Bouet, R. \u0026amp; Lartizien, C. Regularized siamese neural network for unsupervised outlier detection on brain multiparametric magnetic resonance imaging: Application to epilepsy lesion screening. \u003cem\u003eMed. Image Anal.\u003c/em\u003e \u003cstrong\u003e60\u003c/strong\u003e, 101618 (2020).\u003c/li\u003e\n\u003cli\u003eAminpour, A., Ebrahimi, M. \u0026amp; Widjaja, E. Lesion Segmentation in Paediatric Epilepsy Utilizing Deep Learning Approaches. \u003cem\u003eAdv. Artif. Intell. Mach. Learn\u003c/em\u003e \u003cstrong\u003e2\u003c/strong\u003e, 422\u0026ndash;440 (2022).\u003c/li\u003e\n\u003cli\u003eKarthika, A., Subramanian, R. \u0026amp; Karthik, S. Using a recurrent neural network with S2 characteristics, efficient identification of localised cortical dysplasia. \u003cem\u003eJ. Intell. Fuzzy Syst.\u003c/em\u003e \u003cstrong\u003e42\u003c/strong\u003e, 6293\u0026ndash;6306 (2022).\u003c/li\u003e\n\u003cli\u003eSpitzer, H. \u003cem\u003eet al.\u003c/em\u003e Robust and generalisable segmentation of subtle epilepsy-causing lesions: A graph convolutional approach. \u003cem\u003earXiv [eess.IV]\u003c/em\u003e (2023) doi:10.1007/978-3-031-43993-3_41.\u003c/li\u003e\n\u003cli\u003eZhang, X. \u003cem\u003eet al.\u003c/em\u003e Focal cortical dysplasia lesion segmentation using multiscale transformer. \u003cem\u003eInsights Imaging\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 222 (2024).\u003c/li\u003e\n\u003cli\u003eRipart, M. \u003cem\u003eet al.\u003c/em\u003e Detection of epileptogenic focal cortical dysplasia using graph neural networks: A MELD study. \u003cem\u003eJAMA Neurol.\u003c/em\u003e \u003cstrong\u003e82\u003c/strong\u003e, 397\u0026ndash;406 (2025).\u003c/li\u003e\n\u003cli\u003ePastore, L. V. \u003cem\u003eet al.\u003c/em\u003e Integrating standard epilepsy protocol, ASL-perfusion, MP2RAGE/EDGE and the MELD-FCD classifier in the detection of subtle epileptogenic lesions: a 3 Tesla MRI pilot study. \u003cem\u003eNeuroradiology\u003c/em\u003e \u003cstrong\u003e67\u003c/strong\u003e, 665\u0026ndash;675 (2025).\u003c/li\u003e\n\u003cli\u003eHuang, J., Xu, J., Kang, L. \u0026amp; Zhang, T. Identifying epilepsy based on deep learning using DKI images. \u003cem\u003eFront. Hum. Neurosci.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 590815 (2020).\u003c/li\u003e\n\u003cli\u003eZhao, B. \u003cem\u003eet al.\u003c/em\u003e Integrated automatic detection, classification and imaging of high frequency oscillations with stereoelectroencephalography. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 546 (2020).\u003c/li\u003e\n\u003cli\u003eGupta, S., Meena, J. \u0026amp; Gupta, O. Neural network based epileptic EEG detection and classification. \u003cem\u003eADCAIJ\u003c/em\u003e \u003cstrong\u003eabs/2111.03268\u003c/strong\u003e, (2020).\u003c/li\u003e\n\u003cli\u003eVorndran, J. \u003cem\u003eet al.\u003c/em\u003e A deep learning-based histopathology classifier for Focal Cortical Dysplasia. \u003cem\u003eNeural Comput. Appl.\u003c/em\u003e \u003cstrong\u003e35\u003c/strong\u003e, 12775\u0026ndash;12792 (2023).\u003c/li\u003e\n\u003cli\u003eHosseini, M.-P., Tran, T. X., Pompili, D., Elisevich, K. \u0026amp; Soltanian-Zadeh, H. Multimodal data analysis of epileptic EEG and rs-fMRI via deep learning and edge computing. \u003cem\u003eArtif. Intell. Med.\u003c/em\u003e \u003cstrong\u003e104\u003c/strong\u003e, 101813 (2020).\u003c/li\u003e\n\u003cli\u003eLu, Y. \u003cem\u003eet al.\u003c/em\u003e A novel fusion mechanism for multimodal neuroimaging of focal cortical dysplasias. in \u003cem\u003eProceedings of the 2023 10th International Conference on Biomedical and Bioinformatics Engineering\u003c/em\u003e 28\u0026ndash;37 (ACM, New York, NY, USA, 2023).\u003c/li\u003e\n\u003cli\u003eZhao, X. \u003cem\u003eet al.\u003c/em\u003e Deep learning approaches for imaging-based automated segmentation of tuberous sclerosis complex. \u003cem\u003eJ. Clin. Med.\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, 680 (2024).\u003c/li\u003e\n\u003cli\u003eXu, H., Dong, M., Nakai, Y., Asano, E. \u0026amp; Jeong, J.-W. Automatic detection of eloquent axonal pathways in diffusion tractography using intracanial electrical stimulation mapping and convolutional neural networks. in \u003cem\u003e2018 IEEE 15th International Symposium on Biomedical Imaging (ISBI 2018)\u003c/em\u003e 1034\u0026ndash;1037 (IEEE, 2018).\u003c/li\u003e\n\u003cli\u003eXu, H. \u003cem\u003eet al.\u003c/em\u003e Objective detection of eloquent axonal pathways to minimize postoperative deficits in pediatric epilepsy surgery using diffusion tractography and convolutional neural networks. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e \u003cstrong\u003e38\u003c/strong\u003e, 1910\u0026ndash;1922 (2019).\u003c/li\u003e\n\u003cli\u003eGruen, J., Bauer, T., R\u0026uuml;ber, T. \u0026amp; Schultz, T. Deep learning based tractography with TractSeg in patients with hemispherotomy: Evaluation and refinement. \u003cem\u003eNeuroImage Clin.\u003c/em\u003e \u003cstrong\u003e45\u003c/strong\u003e, 103738 (2025).\u003c/li\u003e\n\u003cli\u003eAhmedt-Aristizabal, D. \u003cem\u003eet al.\u003c/em\u003e A hierarchical multimodal system for motion analysis in patients with epilepsy. \u003cem\u003eEpilepsy Behav.\u003c/em\u003e \u003cstrong\u003e87\u003c/strong\u003e, 46\u0026ndash;58 (2018).\u003c/li\u003e\n\u003cli\u003eAhmedt-Aristizabal, D. \u003cem\u003eet al.\u003c/em\u003e Motion signatures for the analysis of seizure evolution in epilepsy. \u003cem\u003eAnnu. Int. Conf. IEEE Eng. Med. Biol. Soc.\u003c/em\u003e \u003cstrong\u003e2019\u003c/strong\u003e, 2099\u0026ndash;2105 (2019).\u003c/li\u003e\n\u003cli\u003eRaviPrakash, H. \u003cem\u003eet al.\u003c/em\u003e Deep learning provides exceptional accuracy to ECoG-based functional language mapping for epilepsy surgery. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e14\u003c/strong\u003e, 409 (2020).\u003c/li\u003e\n\u003cli\u003eRadha, K. \u0026amp; Narayana, S. DirichNet model for detection of TMS-induced speech errors in patients undergoing epilepsy surgery. in \u003cem\u003eICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)\u003c/em\u003e 1\u0026ndash;5 (IEEE, 2025).\u003c/li\u003e\n\u003cli\u003eHossain, M. S., Amin, S. U., Alsulaiman, M. \u0026amp; Muhammad, G. Applying deep learning for epilepsy seizure detection and brain mapping visualization. \u003cem\u003eACM Trans. Multimed. Comput. Commun. Appl.\u003c/em\u003e \u003cstrong\u003e15\u003c/strong\u003e, 1\u0026ndash;17 (2019).\u003c/li\u003e\n\u003cli\u003eZhang, Y. \u003cem\u003eet al.\u003c/em\u003e Characterizing physiological high-frequency oscillations using deep learning. \u003cem\u003eJ. Neural Eng.\u003c/em\u003e \u003cstrong\u003e19\u003c/strong\u003e, 066027 (2022).\u003c/li\u003e\n\u003cli\u003eMo, J. \u003cem\u003eet al.\u003c/em\u003e Automated detection and surgical planning for focal cortical dysplasia with multicenter validation. \u003cem\u003eNeurosurgery\u003c/em\u003e \u003cstrong\u003e91\u003c/strong\u003e, 799\u0026ndash;807 (2022).\u003c/li\u003e\n\u003cli\u003ePark, D. K. \u003cem\u003eet al.\u003c/em\u003e Convolutional neural network-aided tuber segmentation in tuberous sclerosis complex patients correlates with electroencephalogram. \u003cem\u003eEpilepsia\u003c/em\u003e \u003cstrong\u003e63\u003c/strong\u003e, 1530\u0026ndash;1541 (2022).\u003c/li\u003e\n\u003cli\u003eHu, M. \u003cem\u003eet al.\u003c/em\u003e E2SGAN: EEG-to-SEEG translation with generative adversarial networks. \u003cem\u003eFront. Neurosci.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 971829 (2022).\u003c/li\u003e\n\u003cli\u003eGleichgerrcht, E. \u003cem\u003eet al.\u003c/em\u003e Deep learning applied to whole-brain connectome to determine seizure control after epilepsy surgery. \u003cem\u003eEpilepsia\u003c/em\u003e \u003cstrong\u003e59\u003c/strong\u003e, 1643\u0026ndash;1654 (2018).\u003c/li\u003e\n\u003cli\u003eP\u0026eacute;rez-Garc\u0026iacute;a, F. \u003cem\u003eet al.\u003c/em\u003e Simulation of brain resection for cavity segmentation using self-supervised and semi-supervised learning. in \u003cem\u003eLecture Notes in Computer Science\u003c/em\u003e 115\u0026ndash;125 (Springer International Publishing, Cham, 2020).\u003c/li\u003e\n\u003cli\u003eP\u0026eacute;rez-Garc\u0026iacute;a, F. \u003cem\u003eet al.\u003c/em\u003e A self-supervised learning strategy for postoperative brain cavity segmentation simulating resections. \u003cem\u003eInt. J. Comput. Assist. Radiol. Surg.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 1653\u0026ndash;1661 (2021).\u003c/li\u003e\n\u003cli\u003ePantovic, A., Ren, X., Wemmert, C., Ollivier, I. \u0026amp; Essert, C. Generation of synthetic training data for SEEG electrodes segmentation. \u003cem\u003eInt. J. Comput. Assist. Radiol. Surg.\u003c/em\u003e \u003cstrong\u003e17\u003c/strong\u003e, 937\u0026ndash;943 (2022).\u003c/li\u003e\n\u003cli\u003eSu, D. \u003cem\u003eet al.\u003c/em\u003e Automatic, deep-learning-based segmentation of the amygdalohippocampectomy resection cavity in MR images. in \u003cem\u003eMedical Imaging 2023: Image-Guided Procedures, Robotic Interventions, and Modeling\u003c/em\u003e (eds. Linte, C. A. \u0026amp; Siewerdsen, J. H.) 33 (SPIE, 2023).\u003c/li\u003e\n\u003cli\u003eCourtney, M. R. \u003cem\u003eet al.\u003c/em\u003e Automated segmentation of epilepsy surgical resection cavities: Comparison of four methods to manual segmentation. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e296\u003c/strong\u003e, 120682 (2024).\u003c/li\u003e\n\u003cli\u003ePantovic, A., Ollivier, I. \u0026amp; Essert, C. 2D and 3D-UNet for segmentation of SEEG electrode contacts on post-operative CT scans. in \u003cem\u003eMedical Imaging 2022: Image-Guided Procedures, Robotic Interventions, and Modeling\u003c/em\u003e (eds. Linte, C. A. \u0026amp; Siewerdsen, J. H.) (SPIE, 2022). doi:10.1117/12.2606538.\u003c/li\u003e\n\u003cli\u003eBanerjee, S. \u003cem\u003eet al.\u003c/em\u003e Deep relational reasoning for the prediction of language impairment and postoperative seizure outcome using preoperative DWI connectome data of children with focal epilepsy. \u003cem\u003eIEEE Trans. Med. Imaging\u003c/em\u003e \u003cstrong\u003e40\u003c/strong\u003e, 793\u0026ndash;804 (2021).\u003c/li\u003e\n\u003cli\u003eZhang, W. \u003cem\u003eet al.\u003c/em\u003e Matpr-Unet: A Multi Attention Two-Path Residual Unet for Focal Cortical Dysplasia Lesions Segmentation. in \u003cem\u003eICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)\u003c/em\u003e 1641\u0026ndash;1645 (IEEE, 2024).\u003c/li\u003e\n\u003cli\u003eJohnson, G. W. \u003cem\u003eet al.\u003c/em\u003e Localizing seizure onset zones in surgical epilepsy with neurostimulation deep learning. \u003cem\u003eJournal of neurosurgery\u003c/em\u003e \u003cstrong\u003e138\u003c/strong\u003e, 1002\u0026ndash;1007 (2022).\u003c/li\u003e\n\u003cli\u003eFı\u0026ccedil;ıcı, C., Telatar, Z., Ko\u0026ccedil;ak, O. \u0026amp; Eroğul, O. Identification of TLE focus from EEG signals by using deep learning approach. \u003cem\u003eDiagnostics (Basel)\u003c/em\u003e \u003cstrong\u003e13\u003c/strong\u003e, (2023).\u003c/li\u003e\n\u003cli\u003eSharma, R. Localization of epileptic surgical area using automated hybrid approach based on higher-order statistics with sensitivity analysis and residual wavelet transform. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e \u003cstrong\u003e86\u003c/strong\u003e, 105192 (2023).\u003c/li\u003e\n\u003cli\u003eWang, X., Wu, D. \u0026amp; Yang, C. Localization of epileptic foci from intracranial EEG using the GRU-GC algorithm. \u003cem\u003eBrain Inform.\u003c/em\u003e \u003cstrong\u003e12\u003c/strong\u003e, 6 (2025).\u003c/li\u003e\n\u003cli\u003eZhang, Z., Luo, Q. \u0026amp; Zhou, Y. Localization of focal intractable epileptic focus based on long-term recurrent convolutional network. in \u003cem\u003e2023 IEEE 6th International Conference on Pattern Recognition and Artificial Intelligence (PRAI)\u003c/em\u003e 570\u0026ndash;574 (IEEE, 2023).\u003c/li\u003e\n\u003cli\u003eWang, Y. \u003cem\u003eet al.\u003c/em\u003e SEEG-Net: An explainable and deep learning-based cross-subject pathological activity detection method for drug-resistant epilepsy. \u003cem\u003eComput. Biol. Med.\u003c/em\u003e \u003cstrong\u003e148\u003c/strong\u003e, 105703 (2022).\u003c/li\u003e\n\u003cli\u003eDou, Y. \u003cem\u003eet al.\u003c/em\u003e Identification of epileptic networks with graph convolutional network incorporating oscillatory activities and evoked synaptic responses. \u003cem\u003eNeuroimage\u003c/em\u003e \u003cstrong\u003e284\u003c/strong\u003e, 120439 (2023).\u003c/li\u003e\n\u003cli\u003eChen, W. \u003cem\u003eet al.\u003c/em\u003e Unsupervised detection of high-frequency oscillations in intracranial electroencephalogram: promoting a valuable automated diagnostic tool for epilepsy. \u003cem\u003eFront. Neurol.\u003c/em\u003e \u003cstrong\u003e16\u003c/strong\u003e, 1455613 (2025).\u003c/li\u003e\n\u003cli\u003eNorris, J. \u003cem\u003eet al.\u003c/em\u003e Localising the Seizure Onset Zone from Single-Pulse Electrical Stimulation Responses with a CNN Transformer. in \u003cem\u003eProceedings of the 9th Machine Learning for Healthcare Conference: PMLR\u003c/em\u003e (eds. Deshpande, K. et al.) vol. 252 1\u0026ndash;21 (PMLR, Toronto, Canada, 2024).\u003c/li\u003e\n\u003cli\u003eLi, S., Wang, Y., Yang, Y., Wei, P. \u0026amp; Kang, G. CRM: An automatic label generation method based on semi-supervised learning for high frequency oscillatory. in \u003cem\u003eProceedings of the 2022 9th International Conference on Biomedical and Bioinformatics Engineering\u003c/em\u003e 100\u0026ndash;107 (ACM, New York, NY, USA, 2022).\u003c/li\u003e\n\u003cli\u003eLiu, X., Hu, L., Wang, S. \u0026amp; Shen, J. Localization of seizure onset zone with epilepsy propagation networks based on graph convolutional network. \u003cem\u003eBiomed. Signal Process. Control\u003c/em\u003e \u003cstrong\u003e74\u003c/strong\u003e, 103489 (2022).\u003c/li\u003e\n\u003cli\u003eMonsoor, T. \u003cem\u003eet al.\u003c/em\u003e Optimizing detection and deep learning-based classification of pathological high-frequency oscillations in epilepsy. \u003cem\u003eClin. Neurophysiol.\u003c/em\u003e \u003cstrong\u003e154\u003c/strong\u003e, 129\u0026ndash;140 (2023).\u003c/li\u003e\n\u003cli\u003eSadek, Z., Hadriche, A. \u0026amp; Jmail, N. Clustering of high frequency oscillations HFO in epilepsy using pretrained neural networks. in \u003cem\u003eLecture Notes in Networks and Systems\u003c/em\u003e 100\u0026ndash;107 (Springer Nature Switzerland, Cham, 2023).\u003c/li\u003e\n\u003cli\u003eLiu, X. \u003cem\u003eet al.\u003c/em\u003e Focal EEG recognition combining LMD and deep reinforcement learning. in \u003cem\u003eProceedings of the 2023 3rd International Conference on Bioinformatics and Intelligent Computing\u003c/em\u003e 165\u0026ndash;168 (ACM, New York, NY, USA, 2023).\u003c/li\u003e\n\u003cli\u003eGardy, L. \u003cem\u003eet al.\u003c/em\u003e Detecting fast-ripples on both micro- and macro-electrodes in epilepsy: A wavelet-based CNN detector. \u003cem\u003eJ. Neurosci. Methods\u003c/em\u003e \u003cstrong\u003e415\u003c/strong\u003e, 110350 (2025).\u003c/li\u003e\n\u003cli\u003eBroti, N. M. \u003cem\u003eet al.\u003c/em\u003e Multi-modal multitask learning model for simultaneous classification of two epilepsy biomarkers. in \u003cem\u003eLecture Notes in Computer Science\u003c/em\u003e 235\u0026ndash;250 (Springer Nature Switzerland, Cham, 2025).\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":true,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"npj-health-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [npj Health Systems](https://www.nature.com/npjhealthsyst/)","snPcode":"44401","submissionUrl":"https://submission.springernature.com/new-submission/44401/3","title":"npj Health Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false},"keywords":"","lastPublishedDoi":"10.21203/rs.3.rs-8612799/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-8612799/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eArtificial intelligence is increasingly explored to support decision-making in epilepsy surgery, yet evidence for implementation across the epilepsy surgery pathway remains limited. We conducted a scoping review of 145 studies published between January 2018 and May 2025 to map AI-enabled decision support systems across surgical stages and clinical tasks, characterize datasets by modality, size, geographic provenance and accessibility, and synthesize modeling practices, external validation and workflow integration. The literature is heavily concentrated in the pre-operative stage, with no included intra-operative studies and relatively few post-operative applications. Most studies rely on small, single-center and non-public datasets and use supervised CNN-based models. External validation and workflow-integrated evaluation are uncommon, and only a minority of systems report semi-integrated clinical workflows. These findings highlight key gaps in generalizability, workflow readiness and equity, and inform priorities for multi-center data resources, rigorous cross-site evaluation and clinically meaningful endpoints to enable safe, scalable adoption.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003e \u003c/p\u003e","manuscriptTitle":"AI enabled decision support systems in epilepsy surgery a scoping review","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-02-19 09:13:59","doi":"10.21203/rs.3.rs-8612799/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Revision requested","date":"2026-03-11T12:13:56+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-03-08T18:13:57+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"211957465432380446809858736579007906270","date":"2026-02-26T07:32:31+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"146141179345445215879147470145643622615","date":"2026-01-31T12:00:15+00:00","index":"hide","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2026-01-30T17:33:07+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"188972695157569176020070118468991396561","date":"2026-01-26T21:21:11+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2026-01-26T05:28:22+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2026-01-21T00:37:49+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2026-01-19T17:09:23+00:00","index":"","fulltext":""},{"type":"submitted","content":"npj Health Systems","date":"2026-01-15T17:26:10+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"npj-health-systems","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"","sideBox":"Learn more about [npj Health Systems](https://www.nature.com/npjhealthsyst/)","snPcode":"44401","submissionUrl":"https://submission.springernature.com/new-submission/44401/3","title":"npj Health Systems","twitterHandle":"","acdcEnabled":true,"dfaEnabled":true,"editorialSystem":"stoa","reportingPortfolio":"NPJ","inReviewEnabled":true,"inReviewRevisionsEnabled":false}}],"origin":"","ownerIdentity":"81f46f01-2597-41bf-90dc-d06982845d57","owner":[],"postedDate":"February 19th, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"under-review","subjectAreas":[{"id":62980697,"name":"Biological sciences/Computational biology and bioinformatics"},{"id":62980699,"name":"Health sciences/Health care"},{"id":62980700,"name":"Physical sciences/Mathematics and computing"},{"id":62980701,"name":"Health sciences/Medical research"},{"id":62980702,"name":"Scientific community and society/Scientific community"}],"tags":[],"updatedAt":"2026-04-30T21:24:03+00:00","versionOfRecord":[],"versionCreatedAt":"2026-02-19 09:13:59","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-8612799","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-8612799","identity":"rs-8612799","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}