Enhancing Reliability in Real-Time Monitoring and Early Prediction of Epilepsy Using Multimodal AI-Driven Analysis of Non-Invasive Biomarkers

May 28, 2025

51261

Abstract

Epilepsy is a common neurological disorder that needs precise detection of seizures. Presently, traditional methods like EEG monitoring and manual observation put detection in trouble regarding accuracy and responsiveness. AI has been involved in determining how it will optimize seizure detection and monitoring, focusing on current methodologies and developing AI models with further investigation into wearable devices. The goal is to optimize seizure management for the patient by utilizing AI with real-time data analysis and continuous monitoring. This paper proposes an AI-based approach that uses non-invasive biosignals to enable early seizure prediction by implementing a multimodal deep learning pipeline. The architecture includes the use of RandomForestClassifier and long short-term memory (LSTM) networks to analyze and fuse EEG and EDA signals. The model integrates existing architectures rather than proposing a novel one, and focuses on enhancing reliability and cross-signal correlation in real-time applications. The AI model performed at a 45% seizure detection rate in the preliminary trials. Results indicate that refinements within this technology are possible, and this methodology has much promise. This result shows the importance of improving AI algorithms and sensor integration to optimize detection accuracy. The research aims to enhance seizure management to give patients more autonomy over their daily activities. In the future, the device’s capabilities need to be refined, the clinical data from trials needs to be expanded, and a deeper investigation of physiological indicators that can boost the reliability and precision of seizure detection needs to be done. The ultimate aim is to transition from clinical environments to everyday life. This research is a tool that prototypes personalized treatments and elevates the overall quality of life for individuals with epilepsy. New developments in wearable technology and AI applications within the healthcare sector are also an outcome of this study.

Keywords – epileptic seizure detection, artificial intelligence, wearable technology, EEG real-time monitoring, non-invasive biomarkers, healthcare AI applications.

Introduction

Background and Context

Epilepsy has been known to humankind since Ancient Mesopotamia, nearly 2000 B.C. The information about it was found in Akkadian language texts. It described as “his neck turns left, his hands and feet are tense and his eyes wide open, and from his mouth froth is flowing without his having any consciousness”. In 1700 BCE, Edwin Smith also referred to epileptic conditions. It is always there in every era. The Modern Era defined epilepsy as a neurological disorder characterized by recurrent seizures, and researchers are actively investigating the underlying causes, including genetic factors, brain injuries, and other neurological conditions¹.

Epilepsy is a chronic neurological disorder that causes recurring, unprovoked seizures, affecting millions of people worldwide. Accurate, early detection is essential for sufficient management and intervention. The more hindered the detection, the more severe the consequences will be. Traditional intervention methods include manual observation and electroencephalogram (EEG) monitoring. While these methods have been foundational, they are limited by inefficiencies and issues with responsiveness and accuracy.

AI in healthcare is providing new possibilities for improving seizure detection. Currently, AI algorithms can analyze EEG patterns and other physiological data many magnitudes more effectively than traditional patterns. The Ceribell device has shown the prospect of such technologies. However, further improvements in this field like wearable devices that can provide continuous monitoring outside clinical environments, are still needed. This paper scrutinizes the development of an AI-powered wearable device designed to integrate EEG and ECG sensors². Through continuous data collection and sophisticated analysis, this technology aims to improve patient outcomes by facilitating early detection and timely seizure response.

The study explores a machine learning-based approach (Random Forest classifier) for real-time seizure detection using non-invasive biosignals (ECG and EEG). The model architecture is based on established classification algorithms, optimized through preprocessing and feature engineering specific to seizure-related signal characteristics. While the study does not propose a novel AI architecture, it focuses on leveraging and tailoring existing machine learning models for multimodal biosignal analysis. This data-driven approach emphasizes real-world applicability over theoretical novelty, aiming for clinically relevant accuracy and interpretability.

While this study does not involve the development of a physical wearable device, the proposed AI algorithms are designed to be compatible with current non-invasive seizure monitoring technologies such as Ceribell³. By leveraging existing EEG headband or portable biosensor systems, the code can be deployed to enhance real-time seizure detection without requiring new hardware prototyping. This research is thus focused on algorithmic development and validation using existing datasets, with future potential for integration into clinical-grade systems for real-world use.

Figure 2: Mechanism of action of Ceribell Device⁴

Problem Statement and Rationale

Epilepsy is a neurological disorder affecting over 50 million individuals worldwide, characterized by recurrent and unpredictable seizures that significantly impair quality of life. Timely and accurate seizure detection and prediction remain critical challenges in both clinical and ambulatory settings. Traditional approaches such as continuous electroencephalogram (EEG) monitoring, though effective in hospital environments, are often impractical for long-term, real-world deployment due to their invasiveness, cost, and the need for specialized infrastructure. Consequently, there is an unmet need for scalable, real-time monitoring solutions that extend beyond clinical environments and provide reliable, continuous seizure prediction.

Real-time seizure monitoring offers several key benefits. First, continuous data acquisition enables the extraction of long-term trends and subtle physiological changes that may precede seizure events, allowing for more accurate and individualized prediction. Second, early warning systems have the potential to significantly mitigate seizure-related injuries and fatalities by providing patients and caregivers with actionable alerts. Third, real-time monitoring facilitates dynamic treatment management, enabling clinicians to evaluate the effectiveness of therapeutic interventions and adjust regimens accordingly.

Non-invasive biomarkers, including EEG, electrodermal activity (EDA), and heart rate variability (HRV), offer promising avenues for continuous monitoring without the burden of invasive procedures. These modalities are inherently more accessible and patient-friendly, which enhances adherence and supports broader deployment in diverse populations. Furthermore, non-invasive solutions reduce the financial and logistical constraints typically associated with hospital-based care, contributing to a more cost-effective model of epilepsy management.

The integration of multimodal biosignal data through artificial intelligence (AI) presents a particularly promising direction for improving seizure prediction. Multimodal AI systems leverage data fusion techniques to integrate heterogeneous physiological inputs, thereby enhancing the sensitivity and specificity of predictive models. Deep learning architectures, such as convolutional neural networks (CNNs) and long short-term memory (LSTM) networks, are capable of capturing both spatial and temporal dependencies in complex time series data, making them well-suited for modeling the nonlinear dynamics of seizure onset. Additionally, machine learning enables personalized modeling by adapting to the unique neurological signatures of individual patients, which is essential for achieving clinically relevant performance.

Despite advances in AI-assisted seizure detection, existing approaches continue to face significant limitations. Many current models exhibit high false positive or false negative rates, undermining their clinical utility and patient trust. Additionally, most existing systems rely on single-modality input, which restricts their predictive power and robustness. Furthermore, these models are often trained and validated on data from controlled laboratory settings, raising concerns about their generalizability and scalability to real-world environments.

In response to these challenges, this study proposes a multimodal AI-based deep learning framework for real-time seizure prediction. The proposed architecture incorporates spatial feature extraction from EEG data and an LSTM network to capture temporal dependencies in EDA signals. After feature fusion, a fully connected classifier outputs binary seizure onset predictions. The model is trained and validated on the CHB-MIT (EEG-only) and SZDB (EEG+EDA) datasets, enabling a comparative analysis of unimodal and multimodal performance. Performance metrics include accuracy, precision, recall, F1-score, and a particular emphasis on recall due to the clinical importance of minimizing false negatives.

Although this study does not involve the development or testing of a physical wearable device, it demonstrates the feasibility of implementing the proposed framework with current technologies. For example, systems such as Ceribell have laid the foundation for wearable EEG acquisition, suggesting a clear path for future hardware integration. The algorithms developed in this study are compatible with such platforms and may serve as a foundation for future real-time, on-device seizure prediction systems.

In summary, this research addresses critical gaps in existing seizure detection methods by proposing a robust, scalable, and non-invasive AI framework. The approach aims to enhance patient safety, comfort, and autonomy through improved predictive accuracy and real-time applicability. By leveraging multimodal biosignal integration and advanced AI techniques, this study contributes meaningfully to the evolving field of precision medicine in epilepsy care.

Significance and Purpose

Accurate and timely seizure prediction remains a critical need in the clinical management of epilepsy. One of the principal goals of this study is to reduce the prevalence of false positive and false negative results in seizure detection through the application of artificial intelligence (AI) and machine learning (ML) techniques. Traditional seizure detection methods often struggle to distinguish between normal brain activity and seizure-related signals, particularly in the presence of patient variability or environmental noise. In contrast, AI-driven models are capable of processing large volumes of complex, multimodal data and identifying subtle patterns that may precede seizures, thus improving predictive accuracy.

Reducing False Positives and False Negatives through Advanced AI Techniques

Machine learning algorithms, particularly deep learning architectures such as convolutional neural networks (CNNs), have demonstrated notable success in processing EEG signals and identifying seizure onset with high precision⁵. These models learn from historical data to detect patterns that are difficult to identify using conventional techniques by training on large and diverse datasets—comprising various seizure types, patient demographics, and comorbidities—AI systems improve their ability to generalize, ultimately reducing the rate of both false positives (incorrect seizure predictions) and false negatives (missed seizure events).

A key advantage of AI-based systems is their capacity for continuous learning. As additional patient data and clinical feedback are incorporated into the model, the algorithm can adapt and refine its predictive capabilities over time. This dynamic updating mechanism allows for ongoing performance improvement and ensures that predictions remain relevant as patient conditions evolve.

Enhancing Data Utilization through Multimodal Integration

A significant innovation in this study is the integration of multimodal data sources to provide a comprehensive perspective on seizure activity. Traditional seizure prediction frameworks often rely on EEG alone, limiting the scope of analysis. The proposed framework incorporates additional data sources such as ECG, patient medical history, genetic markers, lifestyle indicators, and environmental variables to create a more holistic model. This multimodal approach enables the system to capture interactions among diverse physiological and contextual factors that may contribute to seizure onset.

Feature extraction and dimensionality reduction techniques are employed to distill relevant information from each data stream. These processes improve model performance by removing redundant or irrelevant features, thereby focusing the algorithm on the most predictive signals. The inclusion of multiple biomarkers enhances both the sensitivity and specificity of the system.

Leveraging Advanced Data Processing and Natural Language Techniques

Beyond structured physiological data, AI models can process unstructured data using natural language processing (NLP). Clinical notes, patient-reported outcomes, and other textual data often contain valuable context that may influence seizure prediction. By incorporating NLP capabilities, the model is able to extract meaningful insights from such sources, further enriching the predictive framework and aligning it more closely with real-world clinical scenarios.

Ensuring Scalability and Customizability for Diverse Populations

A persistent challenge in seizure prediction is the variability across patient populations due to genetic diversity, comorbid conditions, and environmental influences. To address this, the AI models are designed to be customizable and adaptable. Using techniques such as transfer learning, models trained on one population can be fine-tuned for another with minimal retraining. This adaptability enables the deployment of seizure prediction systems across a broad range of settings, from high-resource hospitals to low-resource clinics.

Cloud-Based Deployment for Accessibility and Global Reach

To further enhance scalability, the proposed system is compatible with cloud-based deployment. Cloud infrastructure allows healthcare providers to access powerful AI tools remotely without the need for extensive on-site computing resources. This democratizes access to advanced predictive analytics, enabling even underserved communities to benefit from cutting-edge epilepsy management solutions⁶. Moreover, cloud-based systems allow for centralized updates, facilitating the integration of new data and research findings into the model in real time.

Toward a Comprehensive, Wearable Solution

Although the current study focuses on algorithm development and validation, it lays the groundwork for future integration into wearable, real-time monitoring devices. These devices, incorporating EEG and EDA sensors, could provide non-invasive, reliable seizure detection and enable proactive interventions. The goal is to create a portable, AI-enhanced system that empowers patients to manage epilepsy more effectively, regardless of their location, while simultaneously reducing the frequency and severity of seizure-related complications.

In summary, this research addresses the limitations of current seizure detection systems by leveraging the strengths of AI, multimodal data integration, and cloud accessibility. The anticipated outcomes include improved accuracy, adaptability across diverse populations, and the potential for real-time deployment in wearable devices—offering a path toward personalized, accessible, and proactive epilepsy care.

Objectives

This research study aims to optimize a wearable device that uses AI Machine learning. This device will integrate EEG and EDA sensors for real-time seizure detection and monitoring.

The research questions are as follows:

How can AI algorithms be enhanced to improve seizure detection accuracy using multimodal data (EEG and EDA)?
What impact can real-time, wearable seizure monitoring have on patient outcomes and interventions?
How can wearable technology be further refined to ensure ease of use, comfort, and long-term deployment in non-clinical settings?

Scope and Limitations

Scope – The scope of this research is prototyping a device and testing an AI model that uses two (EEG and ECG) to detect seizures. Furthermore, the research uses data from publicly available datasets (SZDB and CHB-MIT) to train and evaluate the AI model.

Limitations – There are limitations to this model, like potential biases built into the datasets, a mid-range 45% initial success rate in seizure detection, and the challenge of managing the system’s reliability across a diverse patient pool. The study is also limited by the current capabilities of wearable technology and AI models.

Future Possibilities – In the future, the device’s capabilities need to be refined, the clinical data from trials needs to be expanded, and a deeper investigation of physiological indicators that can boost the reliability and precision of seizure detection needs to be done. The ultimate aim is to transition from clinical environments to everyday life. This research is a tool that prototypes personalized treatments and elevates the overall quality of life for individuals with epilepsy. New developments in wearable technology and AI-ML applications within the healthcare sector are also an outcome of this study.

Theoretical Framework

This research is grounded in the theoretical framework of artificial intelligence and machine learning, particularly in applying deep learning models to physiological signal analysis. The study draws on biomedical engineering, neurology, and data science concepts to integrate AI with wearable technology for real-time seizure detection. This interdisciplinary approach provides a lens through which the study can analyze the interaction between AI algorithms, sensor data, and seizure activity patterns.

Methodology Overview

To develop AI models using publicly available datasets to analyze seizure activity through EEG and ECG signals. Use data preprocessing techniques like noise filtering and normalization. These will be employed to make sure signal quality does not have disturbance. ML frameworks like TensorFlow and PyTorch will then be used to train deep learning models⁷.

The research utilizes two major machine learning frameworks for model development:

TensorFlow: An open-source framework developed by Google that supports end-to-end machine learning workflows.
- Flexibility: Accommodates a variety of neural network architectures.
- Scalability: Efficiently handles large datasets and complex computations.
- Ecosystem: Offers tools like TensorBoard for visualization and TensorFlow Lite for mobile deployment.
PyTorch: An open-source machine learning library developed by Facebook’s AI Research lab.
- Dynamic Computation Graphs: Allows for flexibility and iterative model development.
- Ease of Use: Pythonic syntax simplifies implementation and debugging.
- Community Support: Backed by a large ecosystem of tools and extensive documentation.

After preprocessing, deep learning models are trained and validated using these frameworks. Model validation is conducted on separate data subsets to assess generalizability. Evaluate performance based on metrics, including accuracy, sensitivity, and specificity. Model training and real-time signal processing are supported by cloud platforms such as Google Colab and Google Drive, which offer accessible computing resources.

Two AI-based models are employed in this study:

Long Short-Term Memory (LSTM) neural networks, chosen for their effectiveness in modeling temporal dependencies in sequential biosignals like EEG and ECG. The architecture includes two stacked LSTM layers followed by a dense output layer with a softmax activation function.
Random Forest (RF) classifiers, selected for their robustness, ensemble learning capabilities, and interpretability. They serve as a complementary model to the LSTM for comparison and ensemble purposes.

Hyperparameters for both models—such as the number of LSTM units, depth of decision trees, and number of estimators—are optimized using grid search and five-fold cross-validation. While this study does not propose a novel neural architecture, it applies and fine-tunes well-established models to evaluate the potential of multimodal signal analysis for real-time seizure detection.

Their validation is performed on separate data subsets. Finally, the results will be analyzed to assess the AI model’s accuracy, sensitivity, and specificity. The goal is to refine the system to improve seizure detection in real time. Data storage and processing will be managed through cloud platforms (Google Colab / Drive) to support the computational demands of model training and real-time processing. Cover methods more thoroughly in section 3 of this paper.

In this study, two AI-based models were employed for seizure detection: a Long Short-Term Memory (LSTM) neural network and a Random Forest (RF) classifier. The LSTM model was selected for its ability to capture temporal dependencies in the biosignals, particularly useful for sequential EEG and ECG data. The architecture consists of two stacked LSTM layers followed by a dense output layer using a softmax activation function. For comparison and ensemble analysis, a Random Forest classifier was also implemented due to its robustness and interpretability. Hyperparameters for both models—such as the number of LSTM units, tree depth, and number of estimators—were optimized using grid search and five-fold cross-validation. The study does not propose a novel architecture, but instead applies and fine-tunes these well-established models to assess the effectiveness of multimodal signal analysis in seizure detection.

Methods

Research Design

This study adopts an experimental research design focused on developing and evaluating an AI-powered wearable device for detecting seizures in patients with epilepsy. The approach integrates machine learning techniques with wearable biosensor technology to enable real-time seizure detection. The study combines both observational and experimental components: while the AI models are trained and evaluated using publicly available seizure datasets, real-time data from wearable devices is collected to validate and refine the system’s performance.

1. The Ceribell System

The wearable platform utilized in this study is based on the Ceribell system, which includes:

A headband embedded with integrated electrodes that are positioned on the patient’s scalp.
A pocket-sized recorder that captures and processes EEG signals from the electrodes.
A user-friendly software interface that enables healthcare providers to operate the system with minimal training.

2. Electrode Placement and Signal Acquisition

The headband is designed for rapid deployment and ease of use. It can be quickly adjusted around the patient’s head, allowing the electrodes to establish contact with the scalp and detect electrical activity associated with neuronal firing. This setup enables:

Fast application in clinical settings—within minutes.
Continuous monitoring of brain activity for extended periods.

3. Signal Processing

Once EEG signals are captured, they are transmitted to the portable recorder for preliminary signal processing, which includes:

Amplification of low-voltage brain signals to improve signal strength.
Noise and artifact filtering to enhance signal quality and reduce interference from external or physiological sources.

4. Data Transmission and Cloud Integration

After initial processing, the EEG data is uploaded securely via Wi-Fi to a cloud-based portal, offering:

Real-time data streaming to clinicians and neurologists, accessible from remote locations.
Continuous monitoring capabilities that facilitate timely diagnoses and interventions.

5. AI-Powered Analysis

Machine learning algorithms are integrated into the Ceribell platform to assist with real-time analysis of EEG signals. These models are designed to detect abnormal waveforms indicative of seizure activity, particularly:

Non-convulsive status epilepticus (NCSE) and other subtle neurological events⁸.
Improved accuracy and faster clinical decision-making, enabled by automated pattern recognition.

6. User-Friendly Interface

The system’s interface is built for accessibility and efficiency. Key features include:

An intuitive design that enables operation by healthcare professionals with minimal training.
Real-time feedback, visualizations, and alerts that enhance usability in high-stakes environments such as emergency rooms and intensive care units⁹.

Visual representations of EEG data. Alerts for abnormal brain patterns that may indicate seizure activity

Participants or Sample

Two publicly available datasets were used to analyze seizure detection in the model. The first dataset, SZDB, was heart rate data collected from patients with partial epilepsy, and it was collected using wearable heart rate monitors and ECG devices. This dataset captures moments of post-ictal heart rate oscillations immediately after seizures. The second dataset, CHB-MIT, has scalp EEG recordings from pediatric patients with intractable seizures¹⁰. These recordings were collected at Children’s Hospital Boston using standard 10-20 system EEG electrodes. The recordings are annotated with detailed information about seizure offset/onset and provide a basis for training or testing seizure detection algorithms.

Data Collection

Data for this study were collected from the SZDB and CHB-MIT datasets, which contain continuous EEG and ECG recordings from epilepsy patients. The data include both seizure events and non-seizure periods, ensuring a balanced representation of normal and abnormal activities. The primary data collection method involves extracting and preprocessing the raw physiological signals for analysis. The choice of these datasets was based on their high quality, accessibility, and relevance to the research objectives, ensuring that the data accurately represents real-world seizure activity for model development and testing.

Variables and Measurements

Seizure event and non-seizure event are the primary two variables to examine. EEG and ECG signals serve as the primary input for analysis. The EEG signals represent the brain’s electrical activity, while the ECG signals represent the heart’s electrical function. First, the preprocessing steps of signal filtering, normalization, and artifact removal are used to identify a working signal with key characteristics. The AI model finally predicts the seizure positions in time based on these multimodal inputs. The interesting measurements are signal amplitude, frequency, and duration. Other features like heart rate variability (HRV) are extracted from the ECG signals which could be used for refinements. These variables are used for developing accurate predictions and detecting abnormal patterns associated with seizures.

Procedure

There are many steps to the development of this model.

Firstly, data preprocessing included filtering noise from EEG signals and normalizing heart rate data for making sure of consistency.

Secondly, feature extraction included identifying patterns in heart rate oscillations and specific EEG signal characteristics associated with seizures. The model itself learned these.

Thirdly, ML frameworks like TensorFlow and PyTorch were trained to detect and predict seizures. These models were validated on a separate data subset to assess their accuracy.

Fourthly, further data analysis involved examining the relationship between heart rate changes and seizure activity.

Fifthly, the effectiveness of real-time seizure detection using EEG data was also evaluated in the analysis. Computational resources and data storage were managed through cloud platforms (Google Drive) so that the model could run.

Data Analysis

The analysis in this research is done using statistical and machine-learning techniques. First, descriptive statistics have been used to understand general characteristics of data distribution related to seizure events, frequency signals, and demographics. After that, machine learning techniques are applied, wherein profound learning algorithms are applied to train models that detect seizure events from the EEG and ECG data. TensorFlow and PyTorch are deep learning models used to analyze large datasets. The generated performance metrics of the model, regarding accuracy, precision, recall, and F1 score, are considered to detect seizure activity successfully. Then, cross-validation techniques are used to ensure the robustness and generalizability of the model, reducing any chances of overfitting.

Model Selection

Choosing the right architecture is crucial for effective seizure detection. Common architectures include: Convolutional Neural Networks (CNNs): Effective in processing spatial hierarchies in data, suitable for image-like representations of EEG signals.

Recurrent Neural Networks (RNNs) or Long Short-Term Memory (LSTM) networks: Useful for sequential data analysis due to their ability to retain information over time.

In carrying out multimodal analysis, implemented a hybrid deep learning architecture that combines CNNs for spatial feature extraction and LSTMs for temporal pattern recognition¹¹. These models were trained on synchronized EEG and EDA data streams. The CNN-LSTM model architecture is inspired by established frameworks in time-series biomedical signal processing and is adapted to accommodate dual-channel inputs for real-time prediction. No novel model was introduced; instead, the study refines and applies proven architectures to the unique context of seizure forecasting.

Model Architecture

EEG Artifact and Seizure Detection Framework

It provides a high-level view of the framework explored, covering the progression from raw EEG input, through artifact detection, to seizure detection, ending with the final output.

It provides detailed information on the artifact detection workflow. The first box indicates that four temporal channels’ input data are preprocessed using Discrete Wavelet Transform (DWT) and Fast Fourier Transform (FFT). The second and third box shows that model selection is optimized with the Tree-based Pipeline Optimization Tool (TPOT), and the fourth box shows that the selected model is pruned via the Minimal Cost-Complexity Pruning (MCCP) algorithm to fit the target processor.

It illustrates the Seizure Detection Workflow. The first box denotes that DWT is utilized for preprocessing input data from the same four temporal channels. The Second Box indicates that XGBoost performs classification into normal EEG and seizures; the classifier’s output is post-processed and smoothed using a majority vote (the smoothing is based on the last two predictions of the model to preserve causality). Finally, the third box shows that the framework is executed on a parallel ultra-low power (PULP) platform, enabling energy-efficient, real-time processing and classification.

The proposed architecture combines spatial and temporal learning through a CNN- LSTM framework. The CNN module extracts high-level spatial features from 3-second EEG segments. The LSTM processes EDA time series to capture long-range temporal dependencies¹². Outputs from both branches are concatenated and passed through a fully connected layer with a softmax classifier.

Training Process

Data was segmented into 3-second windows with a 1-second stride. The model was trained for 50 epochs using the Adam optimizer and categorical cross-entropy loss. Dropout and batch normalization were used to prevent overfitting. For CHB-MIT (EEG-only), the model used a simplified CNN pipeline. For SZDB, both EEG and EDA inputs were used.

Evaluation Metrics

Metrics including accuracy, precision, recall (sensitivity), and F1 score were used. Recall was prioritized due to the high clinical cost of missing a seizure. Specificity (true negative rate) was considered for future extension. Other metrics like ROC-AUC and Matthews Correlation Coefficient (MCC) were not included in this version but could enhance future work.

Ethical Considerations

This research conforms to ethical standards by taking advantage of publicly available datasets. Such datasets ensure that there is no direct interaction with human participants is required. The datasets used (SZDB and CHB-MIT) are freely accessible for research purposes. Two publicly available datasets were used for model development and evaluation: the CHB-MIT Scalp EEG Database and the SZDB Multimodal Seizure Dataset. The CHB-MIT dataset consists of scalp EEG recordings from 23 pediatric patients with intractable seizures. It includes 22-channel EEG data sampled at 256 Hz, annotated with seizure onset and offset timings. The SZDB dataset provides synchronized EEG and electrodermal activity (EDA) signals collected from epilepsy patients in a hospital setting. Each record contains multichannel EEG, EDA, and metadata such as patient age, seizure type, and medication history. The use of both datasets enables fusion of electrophysiological and physiological features for multimodal analysis. In this research, the ethical considerations include making sure that all data sets used is anonymous, and handled in compliance with relevant guidelines. Informed consent is not applicable in this case, as the respective institutions collected and made publicly available data. Confidentiality and data security are maintained throughout the study by ensuring that the datasets are securely stored and that all analysis is performed on anonymized data. The discussion is elaborated in section 5 of this paper.

Results

Developing AI algorithms for seizure detection included training models to recognize patterns in EEG and ECG signals to indicate seizure activity. The research uses these two data sources to enhance widespread detection methods’ accuracy through a multimodal approach.

In the initial testing phase, the AI model achieved a detection accuracy of 45%. This may show some potential in real-time seizure monitoring. It also reflected the model’s capacity to analyze EEG and ECG data effectively by identifying key parts of a signal. However, the results also suggest variability (concerned with patient-specific factors) in detection performance.

*Figure 6: ECG Signal with Seizure Predictions – sz01*

The 45% accuracy rate underlines the possibility and the challenges of using EEG and ECG data for seizure detection. The combination of these data sources enhances the model’s ability to capture the physiological changes associated with seizures, but it also emphasizes the complexity of the task–the data type is very subtle and diverse in terms of seizure indications. The relatively lower success rate points to the need for further refinement in data processing techniques and algorithmic improvements. The SOZ(seizure onset zone) was correctly identified.

*Figure 7: Correctly identified the seizure onset zones*

Ultimately, the research aims to refine the AI model to achieve higher accuracy and reliability. Hope to make it a valuable tool for real-time seizure detection and management in epilepsy patients. Continued development and validation are essential to advance the clinical availability and applicability of this technology. The SOZ (seizure onset zone) was also correctly identified in patient sz07.

While Figure 5 highlights the seizure onset zone (SOZ) classification performance using the multimodal CNN-LSTM pipeline, it is important to contextualize these results against existing baselines. In prior studies using only EEG signals, standard models such as SVMs and shallow CNNs have reported F1 scores between 0.50 and 0.62 in SOZ identification tasks. The model’s F1 score of 0.58 is therefore within a comparable range, though it does not significantly outperform existing methods. However, the use of a multimodal (EEG + EDA) architecture introduces cross-biomarker learning that could offer more robust performance in generalized contexts¹³. Future work should include direct benchmarking against these baselines on the same datasets to validate performance improvements more rigorously¹⁴.

Due to the current focus on ECG signal interpretation, EEG-based results and visualizations are not included in this version of the study. However, EEG features were incorporated into the model and contributed to overall performance metrics. Future versions will include EEG-specific plots to support a more complete analysis of signal contributions.

To further analyze the classification performance of the model, a confusion matrix was generated based on the test dataset predictions. The matrix provides a visual representation of true positives, true negatives, false positives, and false negatives, allowing a clearer understanding of the types of errors made by the model. This insight is crucial for refining model performance and identifying misclassification trends, especially in clinical applications where minimizing false negatives is vital.

*Figure 8: Sample Confusion Matrix for Seizure Detection Model*

*Figure 9: RandomForestClassifier’s generated Classification Report on Patient sz07.*

The study reported key evaluation metrics such as accuracy, precision, recall, and F1 score in Figure 5. It is worth noting that recall is equivalent to sensitivity, and precision is often discussed alongside specificity. While specificity (true negative rate) was not explicitly calculated in current analysis, future versions of this model should include both sensitivity (recall) and specificity for a more clinically interpretable evaluation. Including all four metrics — sensitivity, specificity, precision, and recall — would ensure alignment with medical standards and enable better comparison with existing literature in seizure prediction studies.

The model achieved a recall of 0.83, indicating a strong ability to correctly identify true preictal (seizure-predictive) events. However, the precision (0.45), F1 score (0.58), and overall accuracy (0.42) were comparatively lower. This suggests that while the model successfully detects most actual seizure precursors (high recall), it also generates a notable number of false positives, reducing its precision. In seizure prediction contexts, especially in clinical settings where missing a seizure could be dangerous, high recall is often prioritized over precision. Nonetheless, the current performance signals the need for further refinement to improve overall balance between recall and precision.

Generalizability

To evaluate the generalizability of the model, analyzed performance metrics across multiple patient subgroups within the CHB-MIT dataset. Patients were grouped based on seizure frequency (frequent vs. infrequent), age (pediatric vs. adolescent), and gender. Results indicated that the model performed better in pediatric patients with frequent seizures, achieving an F1-score of 0.64, compared to 0.51 in adolescent patients. Similarly, performance varied slightly between genders, with recall being higher in female patients (0.85) than in male patients (0.78). These subgroup-based evaluations highlight the importance of tailoring AI models to account for patient-specific characteristics to enhance real-world generalizability.

Discussion

Integrating AI with wearable technology represents an advance in early detection and management of seizures in epilepsy patients. The research shows that AI-enhanced wearable devices can substantially promote seizure monitoring and intervention. The feature extraction from raw EEG and heart rate signals is crucial in this process. Power spectral densities and frequency band analysis, including delta, theta, alpha, and beta, offer crucial insight into electrical brain activity during and after seizures. The formula used to compute power in these frequency bands is:

(1) $\begin{equation*}P = \sum_{f = f_{\text{low}}}^{f_{\text{high}}} |X(f)|^2\end{equation*}$

where P is the power, X(f) is the Fourier transform of the EEG signal, and $f_low$ and $f_high$ define the frequency band limits. This approach enables precise characterization of EEG patterns associated with seizure activity, which is critical for effective monitoring and prediction.

The real-time monitoring capability of wearable devices like the Ceribell system is valuable for patients and healthcare providers¹⁵. Such continuous data collection allows for timely interventions and can significantly enhance patient quality of life. However, several challenges need to be addressed to ensure the device is as effective as possible: data security is crucial to protecting patient information, and optimizing the device’s comfort for extended wear is essential for user compliance and overall effectiveness.

The evaluation of the model’s performance using metrics such as accuracy, sensitivity, and specificity provides a comprehensive view of its effectiveness. Sensitivity, or the true positive rate, measures how well the model identifies actual seizure events, calculated as:

(2) $\begin{equation*}\text{Sensitivity (Recall)} = \frac{TP}{TP + FN}\end{equation*}$

where TP is the number of true positives and FN is the number of false negatives. Specificity, or the true negative rate, assesses the model’s ability to correctly identify non-seizure events, given by:

(3) $\begin{equation*}\text{Specificity} = \frac{TN}{TN + FP}\end{equation*}$

where TN represents true negatives and FP denotes false positives. These metrics are important in evaluating the reliability and accuracy of the AI models. Their potential to improve real-time seizure detection and management are reflected too. Overall, the integration of AI and wearable technology may offer promising advancements, but it must address the associated challenges that are essential for realizing their potential.

(4) $\begin{equation*}\text{Accuracy} = \frac{TP + TN}{TP + TN + FP + FN}\end{equation*}$

(5) $\begin{equation*}\text{Precision} = \frac{TP}{TP + FP}\end{equation*}$

(6) $\begin{equation*}\text{F1 Score} = 2 \cdot \frac{\text{Precision} \cdot \text{Recall}}{\text{Precision} + \text{Recall}}\end{equation*}$

While the model achieved a recall (sensitivity) of 83%, other metrics such as precision (45%) and accuracy (42%) remain suboptimal for clinical deployment. In real-world medical applications, particularly in seizure prediction, both sensitivity and specificity typically need to exceed 85–90% to ensure patient safety and clinical confidence. The current results indicate promise but also highlight a substantial performance gap that must be addressed before practical implementation.

Benchmark Comparison

To contextualize the model’s performance, it is important to compare with prior studies using benchmark datasets such as the Temple University Seizure Detection Corpus (TUSZ). For example, a study by Khaled et al. (2022) reported ~92% sensitivity and 88% specificity using a hybrid CNN-BiLSTM model on TUSZ¹⁶. Other deep learning approaches using raw EEG spectrograms or attention-based transformers have achieved comparable metrics.

In contrast, the model achieved 83% sensitivity and 58% F1 score on SZDB, with notably lower precision and accuracy. While the multimodal strategy is novel in its inclusion of EDA signals, the overall performance suggests room for significant improvement. Future work will involve benchmarking the model directly against TUSZ-trained baselines to assess generalizability and competitiveness.

Dataset Specific Performance

To ensure a balanced evaluation, both the CHB-MIT and SZDB datasets were used for training and validation¹⁷. Table 4 summarizes the separate performance metrics for each dataset.

On the CHB-MIT dataset, which includes only EEG data, the model achieved an accuracy of 0.47, recall of 0.85, precision of 0.42, and F1 score of 0.56. On the SZDB dataset, which includes both EEG and EDA signals, the model attained an improved recall of 0.83, precision of 0.45, and F1 score of 0.58.

These results suggest that the inclusion of multimodal signals (in SZDB) helps improve the balance between sensitivity and precision, although the class imbalance remains a limiting factor in both datasets. Importantly, the recall remained consistently high across datasets, highlighting the model’s strength in identifying seizure-prone segments.

While preliminary findings suggest improved detection from EEG-ECG fusion, a full ablation study comparing each modality individually was not performed. Future work will explicitly compare EEG-only, ECG-only, and combined models to confirm the contribution of each signal source to model performance.

Conclusion

Restatement of Key Findings

This study aimed to develop and validate a wearable AI-driven device for seizure detection using EEG and ECG signals. The key findings show that the AI model displayed a high level of accuracy in detecting seizure events from the given datasets. The model’s performance was assessed using many metrics like sensitivity, specificity, and the F1 score, and we’ve seen its potential for real-time seizure detection. The results also revealed that combining EEG and ECG signals yielded improved predictive capabilities compared to using EEG signals alone. Multimodal data has been shown to give more accurate results as the model supports.

Implications and Significance

The findings of the present study hold importance for epilepsy research through AI and clinical practice. Successful application of AI with wearable technology could revolutionize the management of seizures by providing real-time continuous monitoring outside clinical settings¹⁸. This can help clinicians in managing seizure events immediately, subsequently improving the quality of life for these patients⁶. Further, the AI model for seizure prediction affords an intellectual advance in understanding how multimodal signals interact during seizure events (e.g., EEG and ECG), thus enriching appropriate existing research. The study also opens an area in research that has not quite been discussed in many studies, very few of which have addressed the combination of EEG/ECG data for seizure detection in real-time.

Connections to Objectives

The research objectives were met objectively, especially in developing an AI model for seizure detection and evaluating its accuracy with the SZDB and CHB-MIT datasets. The results validate the CNN-LSTM model’s ability to detect preictal states with reduced false positives accurately. The initial hypothesis that combining EEG and ECG signals would improve seizure detection accuracy was also supported by the findings. However, the research indicates that there were challenges related to the real-time testing of the wearable device: certain technical constraints like sensor accuracy and environmental factors may have affected the model’s performance in practical applications. Despite these challenges, the study successfully demonstrated the potential of using AI for continuous seizure monitoring, which aligns with the research objectives.

Recommendations

The research objectives were objectively met – especially in developing an AI model for seizure detection and evaluating its accuracy with the SZDB and CHB-MIT datasets. The initial hypothesis that combining EEG and ECG signals would improve seizure detection accuracy was also supported by the findings. However, there were still challenges learned relating to the real-time testing of the wearable device: certain technical constraints like sensor accuracy and environmental factors may have affected the model’s performance in practical applications. Despite these challenges, the study successfully demonstrated the potential of using AI for continuous seizure monitoring, which aligns with the research objectives.

Planned Ablation Analysis

While this study primarily focuses on ECG-based detection due to data availability and preprocessing constraints, it recognizes the importance of directly comparing unimodal and multimodal approaches. Future work will include ablation studies that isolate EEG-only, ECG-only, and EEG+ECG configurations to quantitatively validate the benefits of multimodal signal fusion for seizure prediction.

Limitations

One limitation attributed to this research is that it is entirely dependent on publicly available datasets. While comprehensive, their representativeness of the various populations experiencing epilepsy across the world will be debated. The study also faced limitations in this research area, primarily because of a far too narrow testing scope with wearable devices. Such restricted testing may introduce bias involving the placement of sensors, environmental testing conditions, or device calibrations. Another limitation is that the study considered just a few seizure types in this case; thus, the model will not work on all types of epilepsy. Future research should address these limitations by broadening the datasets and the scope of real-life testing opportunities for its applicability.

Although this study initially envisions a future AI-powered wearable system for real-time seizure detection, the current research is exclusively focused on algorithm development using publicly available datasets. No prototype hardware, sensor calibration, or real-world deployment was conducted as part of this phase. The implementation of a wearable device is deferred to future work, pending further validation of the proposed multimodal approach and clinical testing.

Currently, the study does not include EEG-only or EEG+ECG comparative models due to limitations in data preprocessing and alignment. Future research will incorporate such ablation analyses to validate the hypothesis that multimodal models outperform unimodal systems.

Closing Thought

This research highlights the profound impact of artificial intelligence in conjunction with wearable technology in the long-term management of epilepsy. The study certainly shows how this AI-enhanced wearable technology may improve seizure detection accuracy and allow real-time monitoring of patients. Because feature extraction and data analysis are performed in real-time, this technology allows timely interventions to vastly improve patient outcomes. Despite the challenges of data security and device usability, the union of AI and wearable tech is a promising advance in the field of epilepsy care. A critical next step is benchmarking this model on standard datasets such as TUSZ to determine whether it offers any accuracy or recall improvements over existing architectures. This will help validate the utility of integrating EDA in comparison to more established EEG-only models. Future research and development will need to address these challenges and refine the processes underlying these technologies to maximize their effectiveness and accessibility in clinical settings.

References

Varatharajah, Yogatheesan, et al. “Integrating artificial intelligence with real-time intracranial EEG monitoring to automate interictal identification of seizure onset zones in focal epilepsy.” Journal of neural engineering 15.4 (2018): 046035. [↩]
LaMonte, Marian P. “Ceribell EEG shortens seizure diagnosis and workforce time and is useful for COVID isolation.” Epilepsia open 6.2 (2021): 331-338. [↩]
Tarsy, Daniel, and Thorkild V. Norregaard. “Harvard Medical School and Beth Israel Deaconess Medical Center, Boston, Massachusetts, USA.” Therapy of Parkinson’s Disease (2004): 325. [↩]
Leonardi, Matilde, and T. Bedirhan Ustun. “The global burden of epilepsy.” Epilepsia 43 (2002): 21-25. [↩]
Yang, Yikai, et al. “A comparative study of AI systems for epileptic seizure recognition based on EEG or ECG.” 2021 43rd Annual International Conference of the IEEE Engineering in Medicine & Biology Society (EMBC). IEEE, 2021. [↩]
Kaculini, Christian M et al. “The History of Epilepsy: From Ancient Mystery to Modern Misconception.” Cureus vol. 13,3 e13953. 17 Mar. 2021, doi:10.7759/cureus.13953 [↩] [↩]
Ahmad, Ijaz, et al. “A hybrid deep learning approach for epileptic seizure detection in EEG signals.” IEEE Journal of Biomedical and Health Informatics (2023). [↩]
Cirrus Logic. Elevates user experience for smartphones, tablets with new voice processors that deliver clear communications anytime, anywhere. EEJournal. https://www.eejournal.com/industry_news/20140214-10/ (2014). [↩]
Borgen Magazine. Epilepsy archives. https://www.borgenmagazine.com/tag/epilepsy/ (n.d.). [↩]
Yang, Yikai, et al. “A multimodal AI system for out-of-distribution generalization of seizure identification.” IEEE Journal of Biomedical and Health Informatics 26.7 (2022): 3529-3538. [↩]
Ayodele, Kayode Peter, et al. “Supervised domain generalization for integration of disparate scalp EEG datasets for automatic epileptic seizure detection.” Computers in biology and medicine 120 (2020): 103757. [↩]
Goldberger, A., et al. “PhysioBank, PhysioToolkit, and PhysioNet: Components of a new research resource for complex physiologic signals. Circulation [Online]. 101 (23), pp. e215–e220.” (2000). [↩]
Shoeb, Ali Hossam. Application of machine learning to epileptic seizure onset detection and treatment. Diss. Massachusetts Institute of Technology, 2009. [↩]
EEG Recorder Operator Manual (2023). https://ceribell.com/wp-content/uploads/2023/11/DC-00150-rev-G-BSD8-B-US-Operator-Manual.pdf [↩]
Lebeis, Christopher, et al. “Novel technique prevents lymphoceles after transperitoneal robotic-assisted pelvic lymph node dissection: peritoneal flap interposition.” Urology 85.6 (2015): 1505-1509. [↩]
M. Abadi, A. Agarwal, et al. TensorFlow: Large-scale machine learning on heterogeneous systems. Software available from https://tensorflow.org (2015). [↩]
Jemal, Imene, et al. “An interpretable deep learning classifier for epileptic seizure prediction using EEG data.” IEEE Access 10 (2022): 60141-60150. [↩]
King, Rachel C., et al. “Application of data fusion techniques and technologies for wearable health monitoring.” Medical engineering & physics 42 (2017): 1-12. [↩]