Early Prediction of Drug-Resistant Epilepsy Using Clinical and EEG Features Based on Convolutional Neural Network
Research Background and Purpose
Epilepsy is a spontaneous and severe neurological disorder characterized by recurrent seizures, affecting approximately 50 million people worldwide [1]. Despite recent advances in anti-seizure medications (ASMs), drug-resistant epilepsy (DRE) still affects 20% to 30% of epilepsy patients [1-3]. DRE patients face significant economic, social, and psychological burdens, and it often takes a long time for drug trials to confirm the diagnosis. Early identification of high-risk patients can provide earlier interventions such as epilepsy surgery, neurostimulation, or ketogenic diet therapy.
Previous studies have identified risk factors for DRE, including early onset, high seizure frequency, abnormal electroencephalograms (EEGs), neurological deficits, cognitive impairments, history of trauma, and intracranial structural lesions [5-9]. However, the importance of these factors in newly diagnosed epilepsy patients is still unclear, necessitating comprehensive tools for early identification of high-risk patients.
EEG plays an indispensable role in the field of epilepsy, including diagnosis, treatment, prognosis, and long-term management [12-14]. Some studies have tried to use EEG features to predict ASM treatment outcomes but have mainly focused on visible EEG features such as epileptiform discharges and power spectrums in the δ band. Nonetheless, some invisible EEG parameters such as power, frequency, coherence, and functional connectivity may also be related to prognosis. With the development of deep learning, more and deeper EEG parameters can be extracted to aid in predicting treatment outcomes.
Machine learning, a method that can automatically summarize rules from data and build predictive models, has been widely applied in the medical field in recent years [15]. Deep learning, an important branch of machine learning, can automatically extract various features, handle large-scale data, and achieve accurate classification results [16-17]. Convolutional neural networks (CNNs), a common deep learning method, have been widely used in automatic seizure detection and studies of epileptic discharges [16,18-19]. Currently, traditional machine learning methods combined with feature extraction yield unsatisfactory results in predicting DRE; therefore, this study primarily explores manually extracted features and develops models to predict DRE.
Source and Author Information of the Research Paper
This research paper was jointly completed by Shi-Jun Yang, Shan-Shan Li, Han-Lin Wang, Jin-Lan Li, Cong-Ping Wang, and Qun-Hui Liu from the Departments of Neurology and Medical Ultrasound at Enshi Tujia and Miao Autonomous Prefecture Central Hospital, as well as the School of Medicine at Xi’an Jiaotong University. The paper was published in the 114th issue of “Seizure: European Journal of Epilepsy” in 2024, with the article being available online on December 16, 2023.
Research Methods and Experimental Procedures
Study Subjects and Exclusion Criteria
This retrospective study included newly diagnosed epilepsy patients who received treatment in the Department of Neurology at Enshi Central Hospital from January 2016 to June 2022, totaling 101 patients. Inclusion criteria were newly diagnosed adult epilepsy patients with complete medical records, EEG and MRI examinations conducted before ASM treatment, as well as regular follow-up visits. Exclusion criteria included: patients with epilepsy syndromes, previous neurological diseases, pregnant women, those taking neuroactive medication before EEG, unclear medical records, and poor compliance.
Data Processing and Preprocessing
EEG data used in this study were recorded within 24 hours before starting ASM treatment, with a sampling frequency of 256 Hz. Each patient provided 10 segments of 90-second EEG signals, totaling 1010 segments. In data preprocessing, multiple methods were used to remove artifacts, including removing useless electrodes, filtering, re-referencing, segmenting, baseline correction, discarding bad segments, independent component analysis, and noise elimination. Finally, the original data were converted into one-dimensional “.xls” file format using pandas functions.
Development of Convolutional Neural Network Model
CNN, a common signal and image processing method, was used in this study to automatically extract features from raw EEG signals and refine high-level features. The CNN structure includes convolution layers, a flattening layer, and fully connected layers. Convolution layers automatically extract EEG features through convolution operations; the flattening layer flattens EEG data to improve computation speed and feature robustness; fully connected layers acquire features extracted by the neural network. The activation function used is Rectified Linear Unit (ReLU), and the output function used is Softmax.
System Evaluation and Statistical Analysis
Multiple evaluation metrics were employed to assess model performance, including accuracy, specificity, precision, sensitivity, F1-score, Kappa statistic, mean squared error (MSE), and area under the curve (AUC). The calculation formulas for each metric are as follows:
Accuracy = (TP + TN) / (TP + FN + FP + TN)
Specificity = 1 - (FP / (FP + TN))
Precision = TP / (TP + FP)
Recall = TP / (TP + FN)
F1-score = 2 * (precision * recall) / (precision + recall)
Kappa statistic = (observed accuracy - expected accuracy) / (1 - expected accuracy)
MSE = (1/N) * Σ(observedi - predictedi)
AUC = Σ(rankinsi ∈ positive class) / (m * (m + 1)) - m / n
We developed both EEG models and clinical-EEG models, using a seven-layer architecture including two convolution layers, one flattening layer, and four fully connected layers. Both models were trained for 30 iterations on the training set, yielding final results through testing and validation.
Research Results
In our study, 101 patients were ultimately included (78 with drug-sensitive epilepsy and 23 with DRE). By combining clinical and EEG characteristics to predict DRE in newly diagnosed epilepsy patients, satisfactory results were achieved. In the test set of the EEG model, accuracy, specificity, precision, sensitivity, F1-score, Kappa statistic, optimal MSE, and AUC were 0.99, 0.59, 0.82, 0.90, 0.86, 0.72, 181.76, and 0.76, respectively; in the validation set, accuracy was 0.81. In the test set of the clinical-EEG model, these metrics were 0.99, 0.72, 0.82, 0.96, 0.89, 0.83, 32.00, and 0.81, respectively; in the validation set, accuracy was 0.84.
Significance and Application Value of the Research
This study developed and validated EEG and clinical-EEG models for predicting DRE in newly diagnosed epilepsy patients, providing new tools for early identification of high-risk DRE patients. By combining clinical and EEG data, these models not only improve prediction accuracy but also help clinicians make more targeted treatment decisions at an early stage, avoiding repeated failed ASM trials.
Research Highlights and Innovations
- High-performance predictive model: The models performed excellently, especially the clinical-EEG model, exhibiting superior accuracy, specificity, and precision.
- Wide clinical application prospects: The model utilizes the combination of clinical and EEG features, showcasing high practicality and clinical value.
- Early intervention for newly diagnosed epilepsy patients: By early identification of high-risk patients, alternative treatment methods can be implemented sooner, improving patient quality of life.
Research Summary
This study developed and validated an effective EEG and clinical-EEG model for predicting DRE in newly diagnosed epilepsy patients. This research not only extends the application range of deep learning algorithms in epilepsy prediction but also provides a reliable tool for clinical practice, aiding early identification of high-risk patients, and achieving more precise and personalized treatment.