Epilepsy, a prevalent neurological disorder affecting approximately 50 million individuals globally, is characterized by recurrent, unprovoked seizures that can significantly impair quality of life [1,2]. Effective monitoring and timely detection of seizures are crucial for optimal management and improved patient outcomes, especially for preventing sudden unexpected death in epilepsy [3]. Traditional monitoring methods, such as electroencephalography (EEG), often require wearable devices that can be intrusive and uncomfortable, potentially limiting patient compliance and continuous monitoring capabilities [4]. Recent advancements in artificial intelligence (AI) have paved the way for innovative, nonwearable monitoring systems that use audiovisual data to predict, detect, and warn of epileptic seizures [5]. These systems offer a contactless approach, enhancing patient comfort and enabling continuous monitoring without the physical constraints associated with wearable devices [6]. For instance, vision-based AI systems have demonstrated efficacy in detecting motor seizures by analyzing patient movements in visual data, most commonly video recordings, providing a nonintrusive alternative to traditional methods [7].

The current state of the art in epilepsy monitoring remains largely centered on EEG due to its high accuracy in detecting seizure events [8,9]. EEG-based systems remain the clinical gold standard, offering precise detection capabilities, but they are often resource-intensive because of long setup times, the need for clinical expertise, and the associated long hospital stays [3]. This traditional approach relies heavily on inpatient settings, wearable devices, or induced seizures through medication withdrawal—methods that can be intrusive, resource-intensive, and distressing for patients [10-12]. Given the high costs of hospital stays and long waiting times for diagnostic assessments, home telemonitoring has gained increasing importance [11]. Consequently, there is an emerging demand for innovative, patient-friendly monitoring systems that address these patient needs. For individuals with epilepsy, this involves reducing dependence on hospital-based diagnostic procedures through the implementation of reliable home-based detection systems. This approach is expected to ensure timely interventions, thereby enhancing the general well-being and quality of life of those affected.

A promising development in the field has been the combination of EEG with video-based seizure detection, which integrates the accuracy of EEG with the nonintrusive nature of vision-based monitoring [12]. Advances in deep learning (DL) have further enhanced vision-based seizure detection, improving system performance and providing additional decision support for epilepsy management [13]. These improvements have facilitated a transition from clinical technologies to home-based solutions relying solely on video monitoring, enabling greater accessibility and continuity of care [8]. The ability to capture seizure events in real-world settings further enhances ecological validity and contributes to more personalized treatment approaches [13]. The integration of AI with video data not only alleviates the discomfort associated with wearable devices but also enhances the accuracy of seizure detection. AI algorithms can analyze subtle visual cues and patterns that indicate seizure activity, facilitating early detection and intervention [4,5]. Moreover, nonwearable systems reduce the burden on patients and caregivers, promoting better adherence to monitoring protocols and improving overall quality of life [3].

AI monitoring systems represent a transformative shift in epilepsy detection and management, with vision-based solutions emerging as a key advancement in the field. These systems follow a similar workflow to process video data and analyze it using AI techniques. An illustrative overview of an exemplary schematic flow of monitoring systems based on visual data with AI models is provided in Multimedia Appendix 1. To facilitate a comprehensive classification of the contents of this work, the 5 steps of the process that serve as a contextual foundation are presented below as examples: monitoring, visual data acquisition, visual data preprocessing, AI model processing, and event analysis. At the start of the schematic flow, patient monitoring is initiated using either mobile or stationary systems to enable continuous observation [3,14]. Visual data are acquired through optical sensors that generate video recordings, which are subsequently preprocessed. The goal of preprocessing is to reduce artifacts, that is, unwanted disturbances or distortions that alter the signal, may compromise analytical quality, and are not part of the original input [15]. Common methods include image denoising, video stabilization, and normalization of video data to ensure high-quality input for subsequent AI-based analysis [14]. AI model processing then uses these optimized data, often through DL or convolutional neural networks (CNNs), to identify patterns and estimate seizure probabilities. Preprocessed data are analyzed with neural networks to detect potential clinical events, such as tonic-clonic seizures. Event analysis involves validating these potential events against predefined thresholds and classification criteria. If thresholds are exceeded (eg, event duration or intensity), the events are systematically classified as clinically relevant [16,17]. Automatically generated reports provide detailed information, including the temporal characteristics of the detected events [18]. In addition, a review process may be conducted in which medical professionals validate automated classifications and diagnoses and, when necessary, correct identified events to enhance the accuracy of future AI evaluations [5,18]. Depending on the objective, these monitoring systems either serve seizure detection or support diagnostic processes related to epileptic seizures [5].

The range of applications extends from contactless monitoring of respiratory and heart rates via ultra-wideband or millimeter-wave radar to continuous sleep analysis, reconstruction of mechanical and electrical cardiac activities, and analysis of nocturnal breathing patterns to detect early signs of neurodegenerative diseases, such as Parkinson disease [19]. Moreover, movement sequences are recognized and classified, which is particularly relevant for automatic fall detection in older or care-dependent individuals [20]. Another application includes monitoring correct medication intake, enabling the early identification of potential errors in administration [19]. As vision-based technologies advance, they are anticipated to serve as a fundamental component of future epilepsy monitoring, providing a less invasive and more accessible alternative to traditional methods. While challenges remain, such as ensuring system reliability, addressing environmental variability (eg, fluctuating light conditions or disturbances), and improving user training, ongoing advancements in AI and DL are rapidly enhancing system performance [5,11,13,21,22]. With continuous technological progress, vision-based AI systems have the potential to transform epilepsy care by providing accessible, cost-effective, and user-friendly monitoring in real-world environments.

Previous reviews in this domain have provided valuable but largely isolated insights into specific aspects of AI-based epilepsy monitoring. Their categorizations primarily focused on the scope of monitoring applications [13], target groups such as pediatric or adult patients [3], the period of epilepsy (eg, ictal vs interictal) [3,10], data acquisition sources including EEG and video [23], tracking targets [13], video tracking approaches [21], image processing methods [24,25], and the types of classifiers and performance metrics used [23]. While these contributions have mapped important components of the field, they remain fragmented and mainly descriptive. Whereas earlier reviews have summarized the evidence, our work formalizes it by developing a taxonomy that specifies mutually exclusive and collectively exhaustive dimensions and characteristics. Their initiation occurs on the premise of substantiated proof-of-concept and concrete systems, with refinement occurring through iterative design methodologies. Our taxonomy integrates technical, clinical, and contextual dimensions into a single, coherent schema that enables consistent classification across research and market solutions of AI-enabled video monitoring for epilepsy. By synthesizing these disparate elements into a unified conceptual framework, the taxonomy provides a structured foundation for systematically comparing existing systems, identifying research gaps, and guiding the design and evaluation of future AI-driven monitoring solutions.

Current research leverages diverse data sources and technological methods to advance the field of vision-based AI monitoring [8-10]. Despite prior studies conducting reviews to provide overviews of these approaches [11-14], a structured taxonomy for categorizing these developments remains absent, and little is known about the dimensions and characteristics of AI technologies in vision-based epilepsy monitoring. To address this research gap, this scoping review developed a taxonomy following the method proposed by Nickerson et al [26], which prescribes iterative cycles to identify dimensions and characteristics, and systematically examined AI technologies in vision-based epilepsy monitoring in both research and market technologies. Specifically, this paper clarified the purpose of developing a taxonomy by articulating why it is needed, how it is constructed, and what it contributes based on the extended taxonomy design process (ETDP) [16]. By reviewing the literature and the health care market and incorporating expert insights, it provides a structured overview of existing solutions and their state-of-the-art developments and offers a foundation for future advancements.

This scoping review aimed to develop and validate a comprehensive taxonomy for vision-based AI technologies in epilepsy monitoring. Specifically, we derived core dimensions and defining characteristics from a scoping review and a market analysis, instantiated and refined the taxonomy through iterative design cycles guided by an ETDP, and evaluated its completeness and practical relevance using expert elicitation. In addressing research question (RQ) 1, the taxonomy’s dimensions and characteristics are specified. In addition, RQ2 synthesizes implications and future directions for research and practice. Our final taxonomy serves as a structured framework for classifying AI technologies in vision-based epilepsy monitoring, supporting researchers in organizing the field, developers in creating targeted solutions, and decision-makers in evaluating and selecting appropriate technologies. Furthermore, this taxonomy supports businesses and health care institutions in assessing market-ready systems and regulatory compliance while facilitating a systematic understanding of emerging trends for policymakers.

The following RQs will be addressed:

A taxonomy is built through iterative derivation, instantiation, and evaluation to define the field’s structure, yielding coherent dimensions and characteristics for comparison and decision-making. This work followed an integrated single evidence base with dual outputs design. A scoping review was conducted to provide the empirical basis for both (1) evidence charting and gap mapping and (2) taxonomy development. These complementary outputs pursue a shared objective of structuring and interpreting a heterogeneous, rapidly evolving body of literature at different analytical levels. This design avoids duplication while ensuring methodological coherence between evidence synthesis and framework construction. The scoping review synthesizes the breadth of published work and enables transparent reporting of study characteristics and evidence gaps. Building on this empirical synthesis, the taxonomy translates the identified patterns into a structured, reproducible classification framework that supports consistent comparison across research prototypes and market-ready solutions.

The aim of this scoping review was to systematically classify emerging AI technologies in vision-based epilepsy monitoring through the development of a comprehensive taxonomy. This was accomplished by adopting the methodological framework [26] and incorporating the ETDP [27]. The ETDP facilitates the explicit description of the underlying problem, the observed phenomenon, the target user groups, and the intended purpose. It is organized along the 6 Design Science Research activities—problem identification, objective definition, design and development, demonstration, evaluation, and communication—and introduces iterative entry and exit points to structure cycles of conceptual and empirical work. For evaluation, ETDP adopts the why-how-what framing: clarify the purpose (formative or summative), select the paradigm and techniques (artificial or naturalistic), and specify the evaluated properties, for example, usefulness (Multimedia Appendix 2). Why clarifies function, whether the evaluation is formative or summative [28]. How specifies the environment, if it is an artificial (eg, laboratory studies) or naturalistic evaluation (eg, case studies with real users) and the timing (ex-ante or ex-post) and the methods [28,29]. What focuses on the criteria, including the objective and subjective ending conditions and the defined evaluation criteria [29]. Overall, the integration of ETDP enables rigorous ex-ante and ex-post assessment of taxonomies while maintaining flexibility for evolving phenomena.

The iterative taxonomy development process followed either a conceptual-to-empirical (C2E) or an empirical-to-conceptual (E2C) approach [26]. The deductive C2E method initiated the taxonomy development by systematically analyzing the relevant literature in the field of emerging AI technologies in vision-based epilepsy monitoring. An inductive E2C strategy was used to derive dimensions and characteristics from existing solutions in practice. Multimedia Appendix 3 illustrates the ETDP-aligned process and the 5 iterative cycles conducted. The meta-characteristic of this taxonomy is thus the identification of the defining properties of AI technologies in vision-based epilepsy monitoring. The ending conditions were guided by the objective and subjective criteria [26] (Multimedia Appendix 2). To derive the final taxonomy, a structured 5-iteration development process was conducted, alternating between C2E and E2C approaches. This iterative procedure combined theoretical grounding with practical assessment, with the latter achieved through an analysis of market-ready solutions to ensure relevance to current practice and to confirm practical alignment of this taxonomy (Figure 1).

This methodological approach is particularly suited, given the increasing complexity and heterogeneity of AI- and vision-based monitoring solutions in epilepsy care, which currently lack a unified framework to systematically organize and compare their diverse characteristics. Accordingly, the taxonomy development began with a rigorous articulation of the motivation and necessity for such a structured classification within this domain. The observed phenomenon comprises emerging AI technologies in vision-based epilepsy monitoring, especially those using video data such as video-EEG or stand-alone video-based seizure detection systems, which are playing an increasingly pivotal role in the diagnosis and management of epilepsy. The primary purpose of the taxonomy is to enable a systematic and transparent categorization of these emerging AI technologies, thereby fostering conceptual clarity, facilitating informed clinical and technological decision-making, and uncovering relevant avenues for future research. The taxonomy thus aims to address the needs of health care professionals, researchers, and technology developers by providing a comprehensive overview of existing and emerging solutions. Patients and patient advocacy groups represent an indirect but essential target group, as improved classification ultimately contributes to the development of more tailored and patient-centered technologies.

Combining iterations 4 and 5 within this section is warranted by their shared method (E2C, Delphi study), common participants, and uniform data collection instruments. Both iterations operationalize the application of the taxonomy and its ex-post assessment. The focus shifts from validation of objective to subjective ending conditions. An integrated report better represents the iterative logic of the taxonomy development, ensures comparability across rounds, and provides a coherent audit trail for how the ending conditions were examined.

In the fourth iteration, a second round of the Delphi study [37] was conducted with the same group of experts. As part of this process, the experts received the iteratively developed taxonomy via email, including detailed descriptions of each dimension and characteristic as derived through the previous iterations. They were explicitly instructed to evaluate whether the identified dimensions and characteristics were both necessary and sufficient for classifying AI technologies in vision-based epilepsy monitoring and to determine if any essential aspects were missing, ambiguous, or misclassified. Furthermore, the experts were asked to systematically assess the taxonomy against the taxonomy design principles [26] to examine whether the objective ending conditions were fulfilled.

The fifth iteration involved conducting a third round of the Delphi study [37] with the same group of experts. They reviewed the updated taxonomy and provided structured feedback, particularly with regard to the subjective ending conditions [27]. These conditions include conciseness, comprehensiveness, robustness, explainability, and extensibility. To facilitate structured feedback, experts were provided with a standardized evaluation template that contained explicit questions for each subjective ending condition to elicit qualitative comments and suggestions related to each of these conditions. For conciseness, experts judged whether any dimension or characteristic was redundant or could be merged without loss of meaning. The comprehensiveness of the taxonomy was investigated through its mapping to typical and rare use cases from vision-based epilepsy monitoring to identify potential gaps. Robustness was examined using deliberately atypical vignettes to evaluate whether classification guidance remained reliable under nonstandard conditions. Explainability was evaluated by having experts restate the taxonomy’s distinctions and indicate whether the aspects, dimensions, characteristics, and naming conventions supported coherent interpretation across stakeholders. Extensibility was assessed by requesting proposals for plausible future additions and checking whether these could be accommodated without reworking existing dimensions. Experts then issued an overall judgment on whether the subjective ending conditions were met, with a short rationale. Aligned with the evaluation goal of establishing the taxonomy’s usefulness for health care professionals, researchers, and technology developers, we implemented a summative, ex-post usefulness check within the Delphi study. The same group of experts mapped the taxonomy to typical and edge-case scenarios drawn from vision-based epilepsy monitoring practice and reflected on its applicability for conducting analyses, enabling systematic comparisons, structuring communication across roles, and informing design decisions. Responses were captured within a structured template that elicited brief rationales and judgments of sufficiency [27].

In summary, experts provided a brief overall statement on whether the objective and subjective ending conditions were met, alongside a justification of the usefulness of the taxonomy. The outcomes of these judgments are reported in the Results section.

Reporting the taxonomy through a frequency analysis shows which fields are well served and where the gaps are located. To communicate the state of the field relative to the taxonomy, we conducted a descriptive frequency analysis consistent with step 6 Report Taxonomy of our methodological approach [27]. The unit of analysis was the set of studies included in the scoping review and the results of the market analysis for AI technologies for vision-based epilepsy monitoring solutions. A standardized coding template derived from the final taxonomy guided the extraction of aspect, dimension, and characteristic labels. Using a predefined matrix, 3 researchers independently annotated each study for the presence of characteristics within every characteristic, allowing multilabel coding where the taxonomy permits (eg, multiple data acquisition sources) and otherwise enforcing exactly one characteristic per dimension. Ambiguities and missing information were marked as not reported following explicit decision rules, with disagreements resolved by consensus after calculating interrater agreement. We then aggregated binary indicators to produce per-characteristic frequencies at the level of studies and, where available, at the level of distinct systems. The resulting counts were summarized in tabular form and visualized to communicate coverage across dimensions and characteristics.

This work did not involve medical products or medicinal trials. It combines a Delphi study and a scoping review. According to the guidelines of the research committee of the University of Osnabrück (Germany), scoping reviews do not require ethical approval. Furthermore, the Research Ethics Committee of the University of Osnabrück (Germany) determined that the Delphi study does not constitute human subjects research (decision/date: NHSR/19.09.2025). This Delphi study focused only on professional opinions from adult participants and involved anonymous expert elicitation and no collection of identifiable private information. Therefore, formal ethics approval was not required. Nevertheless, all panelists received written information about the study purpose (including aims, procedures, risks, and data handling) and provided written informed agreement to participate. Participation was voluntary and could be discontinued at any time without penalty. No vulnerable populations were included. All Delphi panelists were adults (aged ≥18 y) and participated voluntarily after receiving information about the study’s aims, procedures, risks, and data handling. No clinical information or identifiable patient data were collected. Expert feedback was documented using coded identifiers and analyzed in aggregate. No direct personal identifiers are reported. All files were kept on a secure institutional platform of the University Osnabrück with access restricted to the study team with password-protected login and handled in compliance with the General Data Protection Regulation. Participants received no financial or material compensation for taking part in this work.

To ensure a transparent and comprehensible presentation of the taxonomy development process, this section provides an overview of the 5 iterative cycles undertaken to derive the final taxonomy. Each iteration contributed incrementally by identifying, refining, or validating dimensions and characteristics, thereby systematically constructing a robust framework for classifying AI technologies in vision-based epilepsy monitoring.

Developed through an iterative methodology, the final taxonomy consists of 23 dimensions and 102 characteristics, capturing the complexity and heterogeneity of current and emerging solutions in this field. In Multimedia Appendix 3, we provide a schematic overview of the iterations and the dimensions identified at each stage, illustrating the evolution of the taxonomy and the structured expansion of its scope.

During the first iteration, we initially identified 16 dimensions: scope [40], target group [22], period of epilepsy [41,42], data acquisition source [43,44], tracking target [45], video tracking [46,47], image processing [14], type of classifier [42,48], performance metrics [41,49], environment [50], seizure classification [51], medical device [17], salient attributes [5], data privacy [18], user interface [45], and user interaction [3]. These newly uncovered dimensions broadened the taxonomy to better reflect the diverse technical and contextual facets of AI-enabled visual monitoring solutions. As the predefined ending conditions [26] had not yet been fulfilled, iteration 2 used an E2C approach, leveraging a Delphi study to integrate expertise from researchers and practitioners. This strategy led to the addition of 4 further dimensions: cryptographic measure, communication mode, response type, and information purpose. The structured expert feedback not only ensured practical relevance but also helped maintain methodological rigor aligned with established quality criteria for taxonomy development. Despite these advancements, the ending conditions remained unmet, leading to iteration 3. In this stage, practical alignment was prioritized by conducting a market analysis of commercial solutions using Crunchbase and targeted internet research. As a result, 3 new dimensions were identified: computing paradigm, connection type, and support [24,39]. This practical focus ensured that the taxonomy adequately captured real-world implementation aspects and industry trends. The ending conditions were not yet met; another iteration followed. Finally, iterations 4 and 5 consisted of a second round of the Delphi study with the same panel of domain experts to rigorously evaluate the taxonomy against both the objective and subjective ending conditions in accordance with Nickerson et al [26]. The experts confirmed the taxonomy’s conciseness, robustness, explainability, comprehensiveness, and extensibility, with no additional dimensions proposed.

All 9 experts expressed explicit satisfaction with the taxonomy’s scope and structure, confirming that it was well aligned with both theoretical rigor and practical relevance. Every identified object was examined, each characteristic classified at least one object, no further dimensions or characteristics were introduced, and cell combinations were unique and nonredundant, thereby fulfilling the objective ending conditions in iteration 4 and concluding the taxonomy development process. In line with the Introduction’s stated purpose (why, how, and what), these results complete the formative, ex-ante stage of the ETDP and motivate a summative check of utility.

Building on this confirmation, the same expert group proceeded to appraise whether the taxonomy is concise, comprehensive, robust, explainable, and extensible for its intended use in iteration 5 to evaluate the subjective ending conditions. Panelists reported that the taxonomy is concise, noting an absence of redundant elements while allowing for focused extensions was helpful. Experts judged the taxonomy as comprehensive and the coverage to be largely sufficient for the domain (P1-P3 and R1-R6). One expert described the final taxonomy as “very comprehensive and complete in the categories considered” (R1). Furthermore, experts indicated that guidance remains reliable for nonstandard or edge scenarios, characterizing robustness as good and emphasizing maintained flexibility and precision (P1-P3 and R1-R6). The experts reported that the taxonomy’s structure supports understanding; distinctions were seen as transparent and traceable, and the systematic differentiation was said to improve explainability and provide a comprehensive overview (P1-P3 and R1-R6). It is acknowledged that certain terminology may not be comprehensible to specialists. For instance, medical terminology may not be readily accessible to IT professionals, and conversely, IT terminology may not be readily comprehensible to medical staff. The incorporation of a detailed description or an extended legend accompanied by comprehensive explanations could prove beneficial in this context (R1 and R6). Facing the condition of extensibility, the forward compatibility was consistently affirmed. Experts noted that, due to its modular structure, the taxonomy can be extended by adding elements without rewriting existing ones (P1-P3 and R1-R6). In aggregate, experts assessed the subjective ending conditions as fulfilled, indicating that the taxonomy is both theoretically coherent and practically usable for its intended purposes.

Experts judged the taxonomy to be useful for its intended target group and purpose (P1-P3 and R1-R6). Panelists provided convergent evidence that the taxonomy is practically useful for analysis, comparison, communication, and decisions. They noted that, in clinical settings, the taxonomy functions as a common language for appraising systems and coordinating with nonclinical stakeholders, ultimately aiding selection and configuration. Clinicians reportedly view it as a shared frame that supports systematic comparison of alternatives and clearer stakeholder communication, thereby informing practical choices in deployment and workflow integration. “This taxonomy could facilitate a systematic analysis of the landscape of AI-based epilepsy monitoring and drive targeted research and development in areas where it is most required,” declares one expert (R5). Researchers indicated that the taxonomy enables structured synthesis and comparability of studies, facilitates the identification of gaps, and provides a stable basis for cumulative evidence building. From a clinical standpoint, experts reported that the taxonomy offers a common structure for evaluating candidate systems and coordinating expectations among care teams and technology partners, which, in turn, supports context-appropriate choices (R4). From a research perspective, it was described as enabling consistent coding, comparison, and aggregation of evidence, while surfacing neglected areas for future work (R2 and R4). In development contexts, experts indicated that it makes market and clinical requirements more transparent and helps align design trade-offs with medical preconditions (R4 and R6). From an industry perspective, experts indicated that the taxonomy clarifies domain constraints and medical prerequisites, translating needs into design requirements and product positioning (R4 and R6). One expert highlighted that the clear delineation of categories creates shared understanding and a robust foundation for decision-making (R1), whereas another underscored that such structure can advance work in areas with the greatest unmet need (R2). Overall, experts characterized usefulness as high and emphasized that the taxonomy’s clear demarcations promote shared understanding and provide a dependable decision base (P1-P3 and R1-R6).

Viewed in its entirety, the evaluation examined whether the taxonomy is sufficient, clear, applicable, and extensible for its intended use. Complementing the formative, ex-ante checks applied during construction (objective and subjective ending conditions across iterative C2E and E2C cycles), the Delphi study offered an ex-post assessment by domain experts. Their consensus confirmed that the taxonomy’s scope and structure are well aligned with the field and that no further dimensions were required (P1-P3 and R1-R6). This step closes the ETDP loop regarding the demonstration and evaluation phase. The developed taxonomy is theoretically coherent and practically usable for classifying AI technologies in vision-based epilepsy monitoring and for supporting consistent study design and reporting.

As demonstrated in iteration 3 of the practical assessment, the market analysis revealed that many systems have not yet reached market readiness. Given this constrained evidence base, we included one real-world application as an illustrative use case rather than as the basis for a comparative market analysis. Specifically, we provide a detailed classification example of the market-ready system NELLI (Multimedia Appendix 11) [39]. As the number of real-world applications was insufficient for robust market-level inference, the subsequent synthesis focused on the scoping review evidence. To synthesize insights and highlight gaps, we conducted a frequency analysis across the studies of the scoping review (Multimedia Appendix 7). The analysis included 40 studies with multiple coding permitted per dimension when single studies implemented more than one approach. Therefore, frequencies represent counts of appearances rather than mutually exclusive proportions. The corresponding cases are reported transparently below (Table 3).

The aspect of application and context for AI technologies in vision-based epilepsy monitoring reveals the predominant aims to be detection (n=34, 85%) and classification (n=24, 60%), while prediction is almost absent (n=2, 5%). This pattern underscores a focus on immediate recognition of ictal events consistent with clinical priorities, whereas predictive systems remain an open area, likely reflecting their higher methodological complexity. Multiple entries for detection and classification reflect multiscope designs within the same study [17,18,55,56,62]. Target group analysis shows that the overwhelming majority of technologies are tailored to medical professionals (n=33, 82.5%) and secondly to data scientists (n=17, 42.5%), with only marginal attention given to patients with epilepsy themselves (n=8, 20%) or caregivers (n=2, 5%). This suggests a prevailing research trend toward supporting diagnostic and treatment processes within clinical workflows while largely neglecting patient-centered or caregiver-focused designs. The environmental context predominantly involves stationary settings (n=35, 87.5%) compared to mobile scenarios (n=7, 17.5%), reflecting the common use of video-EEG laboratories and controlled observation environments. Similarly, seizure classification heavily favors motor symptoms (n=37, 92.5%) over nonmotor symptoms (n=8, 20%), likely because motor phenomena are more easily captured via visual data, whereas nonmotor events often require multimodal sensing or subjective reporting. Regarding the period of epilepsy, the ictal phase is the principal focus (n=37, 92.5%), followed by preictal (n=3, 7.5%) and postictal (n=3, 7.5%) phases, with limited attention to interictal periods (n=9). This underlines a current emphasis on real-time seizure monitoring over long-term interictal assessment. The data acquisition landscape shows a diverse application of technologies: depth sensors (n=9, 22.5%), infrared (n=20, 50%), 2D cameras (n=36, 90%), 3D cameras (n=5, 12.5%), and video-EEG (n=21, 52.5%) are all well represented. Multimodal systems based on electrocardiography (n=3, 7.5%) and audio (n=6, 15%) are far less frequent, indicating that more multimodal approaches are necessary. Multiple entries for data sources arise when studies use several modalities concurrently [18,42,56,66,70,71].

The visual analysis also highlights the clear dominance of body-focused tracking (n=30, 75%) and sleeping area targeting (n=23, 57.5%), which likely reflects the typical contexts of nocturnal seizure detection. Multiple entries occur where similar events span body and sleeping area [44,60,70]. Face tracking (n=13, 32.5%) appears as a secondary priority, while room-wide overviews (n=2, 5%) are notably rare. For video tracking operators, most studies use movement dynamics (n=30, 75%), and many adopt appearance or feature-based pipelines (n=31, 77.5%). Movement of interest is reported in 16 studies (n=16, 40%), region of interest is reported in 17 studies (n=17, 42.5%), head movement detection appears in 3 studies (n=3, 7.5%), simple keypoint system in 8 studies (n=8, 20%), and biomechanical characteristics in 1 study (n=1, 2.5%). Multiple entries were made where single studies combined several operators [40,47,54,55]. For image processing, optical flow is comparatively common (n=13, 32.5%), followed by contrast-based analysis (n=11, 27.5%), whereas frame differencing (n=3, 7.5%) and spatiotemporal interest points (n=6, 15%) are used infrequently.

Within the aspect of AI model, CNNs are the most used classifiers (n=22, 55%), closely followed by LSTMs (n=11, 27.5%), with other models such as support vector machine (n=8), inflated 3D (n=4, 10%), multilayer perceptron (n=4, 10%), Gaussian mixture model (n=3, 7.5%), and random forest (n=2, 5%) used considerably less. This trend underscores the preference for DL approaches suited for complex spatiotemporal patterns in video data. In this dimension, multiple entries were acceptable because of multiple approaches in studies within AI models based on, for example, LSTM and CNN [43,44,51,65]. Performance evaluation predominantly relies on sensitivity (n=25, 62.5%), accuracy (n=20, 50%), specificity (n=16, 40%), precision (n=15, 37.5%), area under the curve (n=13, 32.5%), and F₁-score (n=14, 35%), while more nuanced metrics such as recall (n=8, 20%) and false-positive rates (n=12, 30%) are less consistently reported. This points to a need for more comprehensive benchmarking that captures the practical trade-offs between detection sensitivity and false alarm rates.

Market identity is dominated by proof-of-concept solutions (n=34, 85%), with relatively few certified medical devices (n=6, 15%). Reported salient attributes cluster around environmental robustness (n=4, 10%), cost-efficiency (n=4, 10%), real-time analysis (n=3, 7.5%), ease of use (n=4, 10%), and high system performance (n=3, 7.5%). Privacy and security remain underreported. Anonymization (n=7, 17.5%) and pseudonymization (n=2, 5%) are occasionally specified; synthetic data (n=0, 0%) and explicit no privacy preserving measure (n=0, 0%) do not appear. Encryption in transit is sometimes stated (n=3, 7.5%), whereas encryption at rest (n=0, 0%) is not, and no encryption (n=2, 5%) is rarely reported.

Reporting on system architecture is uneven. User interfaces include web platforms (n=6, 15%), mobile apps (n=4, 10%), desktop apps (n=3, 7.5%), and wearables (n=1, 2.5%), with voice assistants (n=0, 0%). User interaction is mainly interactive (n=10, 25%) and reporting (n=6, 15%), less with no interaction (n=1, 2.5%), and adaptive interactions are absent (n=0, 0%). Computing paradigms split between cloud-based (n=7, 17.5%) and local on device (n=14, 35%), with edge-based platform (n=2, 5%). The connection type is rarely specified with Wi-Fi (n=2, 5%), built-in modem (n=1, 2.5%), Ethernet (n=3, 7.5%), and Bluetooth (n=1, 2.5%). Support modalities show system setup (n=3), daily technical checks (n=3, 7.5%), expert data review (n=3, 7.5%), and on call (n=1, 2.5%), but help center and chat are both absent (n=0, 0%).

To complete the results, the feedback system indicates that most implementations are geared toward retrospective review rather than immediate intervention. Communication modes are led by on-demand use (n=5, 12.5%), with periodic and event-based configurations occurring less frequently (each n=2, 5%), and true real-time communication remaining uncommon (n=2, 5%). Response types are dominated by visual and text-based outputs (each n=4, 10%), while auditory responses are rarely reported (n=1, 2.5%) and haptic feedback is absent (n=0, 0%). In terms of purpose, the literature primarily documents performance evaluation (n=4, 10%), with comparatively few systems supporting alerting or warning (n=2, 5%) and none providing recommendation or user-learning functions (each n=0).

This frequency analysis provides a structured overview of the current landscape of vision-based AI technologies for epilepsy monitoring and highlights prevailing research emphases as well as underexplored areas. By quantifying recurring dimensions and characteristics, this analysis complements the taxonomy by offering an additional empirical perspective on where the field currently concentrates its efforts and where evidence remains sparse. Beyond summarizing patterns, the frequency analysis increases the practical utility of the taxonomy by enabling systematic comparison across studies and supporting identification of research gaps. In practical terms, the resulting overview can support multiple stakeholder groups: researchers may use it to classify and contrast approaches and to prioritize future investigations, clinicians and decision-makers may use it to appraise the maturity and fit of available technologies for specific care contexts, and developers or companies may use it to benchmark solution profiles and identify opportunities for product development.

In this scoping review of vision-based AI technologies for epilepsy monitoring, we synthesized the evidence into a comprehensive taxonomy structured around application and context, visual analysis, AI model, system architecture, and feedback system, which was subsequently evaluated and refined through domain-expert input. Mapping the included sources to the taxonomy shows that current work is dominated by seizure detection and classification (prediction remains rare) and is largely situated in stationary, hospital ward settings, with comparatively limited evidence from home or residential contexts. Moreover, the field currently shows limited evaluation maturity: most systems remain proof of concept or pilot stage, while deployed solutions are rarely reported.

Building on these principal findings, we derive a set of implications for research and practice (summarized in Multimedia Appendix 12), aimed at supporting researchers, clinicians, developers, policymakers, and patient advocacy groups.

This scoping review has several limitations. Despite a rigorous approach, the scoping review may still reflect selection bias because it was screened and selected by the reviewers, restricted to 3 bibliographic databases, and relied on predefined keywords and inclusion criteria. The exclusion of studies published in languages other than English has the potential to constitute an additional bias. Although the scoping review was conducted across 3 key databases commonly used in health and AI research, scoping reviews may benefit from an even more extensive search strategy, incorporating additional databases or gray literature, to further enhance methodological completeness. However, expert validation and iterative development ensure its robustness. Furthermore, the iterative E2C methodology inherently relies on interpretative refinements, introducing contextual dependencies. Nonetheless, this approach is well established in taxonomy development, allowing adaptability to technological and clinical advancements. The status quo on market-ready solutions appears limited in the landscape of AI monitoring systems that are based on visual data. NELLI [39] provided a basis for the iteration of practical assessment. To obtain a more comprehensive perspective on additional market solutions in the close field, the focus was directed toward a video-EEG solution known as SEER [24]. The performance metrics within AI models pose another challenge, given the multitude of metrics that can be considered. Furthermore, despite the underreporting of privacy and cryptographic safeguards in the literature, a jurisdiction-specific compliance analysis was not undertaken (eg, AI Act). This work does not include a detailed, jurisdiction-specific compliance mapping, which will be addressed in a dedicated follow-up study. Finally, the validation to date relied on the contributions of domain experts; patient and informal caregiver perspectives, and prospective real-world deployments were not yet included.

This scoping review points to a broader transition problem in vision-based epilepsy monitoring: while algorithmic feasibility for seizure recognition is increasingly demonstrated, translation into deployable, trustworthy systems remains constrained by system-level requirements. Progress toward deployment in this domain will likely depend less on incremental model improvements and more on development and evaluation practices that explicitly address readiness for real-world operation, including robustness under uncontrolled conditions, workflow fit, feedback and escalation design, and governance mechanisms that make privacy protection and safety assurance auditable and maintainable over time.

The developed taxonomy provides a shared reference structure for describing and evaluating vision-based epilepsy monitoring technologies end to end. Its innovation lies in integrating application context with visual analysis, AI modeling, system architecture, and feedback design within a single classification framework. Where prior work has often centered on algorithms, datasets, or modality-specific performance summaries, this taxonomy enables consistent, system-level characterization across research prototypes and emerging solutions. By standardizing how systems are characterized across dimensions and characteristics, it can support benchmarking and implementation-focused evaluation (including procurement considerations) and guide implementation priorities. For implementation stakeholders, the taxonomy can serve as a structured checklist to align a system’s intended use and deployment setting with required architecture, feedback design, evaluation evidence, and safeguards, thereby informing selection, integration, and rollout planning. More broadly, the taxonomy supports progress in a diverse and fast-evolving field by making system boundaries, intended use, and evaluation targets explicit and comparable across studies. This can help shift the literature from model-centered reporting toward system-level evidence that can be interpreted across contexts, strengthening the basis for synthesis, replication, and translation. In this sense, the taxonomy is not only a classification tool but also a practical framework to support more consistent study design, reporting, and deployment-oriented evaluation in future research.