An Enhanced Machine Learning-Based Multimodal Framework for Seizure Detection Using EEG and MRI Data

One of the most common neurological disorders that immediately alters a person's way of life is an epileptic seizure. Accurate seizure detection remains a major challenge in neurological research due to the nonstationary and complex nature of electroencephalogram (EEG) signals. Most of the current seizure detection methods use machine learning (ML) and deep learning (DL) models, which highly depend on EEG signals due to their real-time assessment nature. However, these methods still suffer from several limitations: (1) class imbalance problem and (2) discriminative feature extraction. These challenges lead to poor diagnostic accuracy and reliability. This research proposes an innovative unified framework for automated epileptic seizure detection (ESD), integrating EEG and magnetic resonance imaging (MRI) data through advanced DL techniques. In this work, preprocessing is carried out via two distinct streams; that is, initially, high-frequency noise in the EEG signals is removed by applying a Butterworth filter. Then, a data standardization technique is used to normalize the signal, and then signal conversion is carried out using synchrosqueezing wavelet transform (SWT) to represent the signal in a time-frequency domain. Once the EEG is preprocessed, MRI data are preprocessed using data augmentation techniques that improve the diversity of the dataset. After preprocessing, feature learning is performed separately for both EEG and MRI via the improved activation function-based depthwise convolutional neural network (IADCNN). After that, cross-modal attention fusion (CMAF) is utilized to capture more important features obtained from the MRI and EEG, that is, fused data. After that, the seizure detection is carried out by utilizing the custom loss based on the Extreme Gradient Boosting (CLXGBoost) model. Finally, the confidence calibration strategy is applied to predict the reliability of the proposed framework. Unlike existing hybrid models, the proposed framework integrates feature learning, cross-modal fusion, imbalance-aware classification, and confidence calibration within a single unified learning pipeline. Evaluation is conducted on publicly available CHB-MIT and Neuroimaging Tools and Resources Collaboratory (NITRC) datasets. The system achieved an accuracy of 99.56% under the experimental conditions, illustrating improved performance compared to selective baseline approaches. The outcomes demonstrated that the proposed multimodal approach can improve seizure detection performance effectively under experimental settings considered in this study.