EEG Signal Classification Using Fast Fourier Transform and Convolutional Neural Networks: A Hybrid Deep Learning Approach for Binary Classification of Normal and Abnormal Brain Activity
DOI: Pending DOI
Synopsis
Accurate EEG interpretation is challenging for non-specialists; it requires recognition of frequency-band structure, epileptiform discharges, abnormal slowing, and artefacts, and is subject to inter-rater variability. We developed a hybrid framework combining FFT spectral processing with a 2D CNN to classify EEG recordings on a binary classification. Training and internal testing used the CHB-MIT Scalp EEG Database pediatric and the Siena Scalp EEG Database. Eleven fully de-identified clinical recordings from Aster Royal Al Raffah Hospital, Oman, served as an independent external validation cohort. Delta and theta spectral features were the most discriminative for abnormal classification, consistent with established clinical EEG interpretation. The framework is intended as an interpretable educational screening aid under specialist supervision, not as an autonomous diagnostic device.
Background
EEG is a core diagnostic investigation in clinical neurology for epilepsy, encephalopathy, and disorders of consciousness, but reliable interpretation requires specialist training and is prone to significant inter-rater variability1, 4. Clinical EEG is assessed through characteristic frequency bands — delta, theta, alpha, and beta — with pathological patterns including generalized slowing, epileptiform discharges, and ictal rhythms2, 3. During a clinical shadowing program at Aster Royal Al Raffah Hospital, Muscat, Oman (June–July 2025), nursing staff, paramedics, and junior trainees were consistently observed to struggle with rapid recognition of abnormal EEG patterns, and no accessible automated screening tool was available to guide non-specialists. Commercial EEG platforms are proprietary and are not designed as trainee-facing educational tools, while real-time seizure-detection systems target continuous monitoring rather than user uploads8, 9. CNN-based deep learning has demonstrated strong EEG-classification performance, yet accessible, upload-based educational tools remain rare5, 11, 13. This gap motivated the present work.
Methods
Three data sources were used. The CHB-MIT Scalp EEG Database (PhysioNet) provided pediatric scalp EEG; owing to resource constraints, ten subjects (chb01–chb10) were used, with binary labels assigned from seizure annotations (files containing at least one annotated seizure period = abnormal)16, 24. The Siena Scalp EEG Database (PhysioNet) provided adult recordings, broadening the demographic range, with labels derived from the accompanying seizure annotations25, 26. Eleven fully de-identified clinical recordings from Aster Royal Al Raffah Hospital (five normal, six abnormal; labelled from clinical reports) were reserved exclusively for external validation and were not used for training or hyperparameter selection. A uniform preprocessing pipeline was applied: per-channel z-score normalization, standardization to 23 channels (the first 23 channels selected, or zero-padded if fewer), and truncation/padding to 2,500 samples. FFT was applied per channel and the positive-frequency half (1,250 bins) retained, yielding a 23 × 1,250 spectral feature map per recording6. A 2D CNN was implemented in PyTorch with five convolutional layers (16 → 32 → 64 → 128 → 256 filters; 3 × 3 kernels; ReLU; MaxPool2D after layers 1–4), a 64-unit fully connected layer, and a 2-unit softmax output (approximately 1.2 million parameters)5, 11, 12. Training used the Adam optimizer (learning rate 0.001), class-weighted cross-entropy loss, and a weighted random sampler, for 30 epochs at a batch size of 16. The pooled public datasets were partitioned by a stratified 60/40 record-wise split (random_state = 42). Performance was assessed using accuracy, precision, recall, and F1-score, with F1-score prioritized because class imbalance can make accuracy misleading4, 13.
Results
In an exploratory training run, the model achieved approximately 88% accuracy on the held-out test partition (40% of the pooled CHB-MIT and Siena data, stratified by label), with F1-scores of 0.95 (normal class) and 0.91 (abnormal class). These results derive from a single exploratory run; a verified complete reproducible run had not been produced at the time of submission, owing to computational-resource constraints, and should be interpreted as provisional. External validation on the eleven anonymized hospital recordings yielded approximately 82% accuracy, with nine of eleven recordings correctly classified: all five normal recordings and four of six abnormal recordings. Both misclassified recordings were borderline abnormal cases classified as normal by the model — one showed intermittent theta slowing without frank epileptiform activity, and the other contained prominent movement artefact partially obscuring the pathological pattern. The drop from test-set to external performance (approximately 88% to approximately 82%) is consistent with cross-institution generalization losses reported in the literature9, 15. Spectral analysis indicated that delta (0.5–4 Hz) and theta (4–8 Hz) band power were the most discriminative features for the abnormal class, with abnormal recordings showing consistently elevated low-frequency energy — consistent with clinical frameworks in which generalized slowing reflects encephalopathy, post-ictal states, or active seizure activity2, 3. Acknowledged limitations include a record-wise rather than patient-wise split (which can inflate apparent accuracy), a Siena labeling implementation issue (filename-based rather than annotation-file-based) that introduced label noise on that subset, and an external cohort of eleven recordings that is too small for statistically robust conclusions13. Quantitatively, mean normalized delta power was 81% higher in abnormal recordings than in normal recordings, theta power 68% higher, while alpha and beta power were 31% and 24% lower, respectively.
Discussion
These results support the feasibility of frequency-domain deep learning as an interpretable educational aid for binary EEG screening13, 14. Explicit FFT preprocessing encodes established neurophysiological priors, making model outputs more transparent than end-to-end time-domain classifiers — a property as important as raw accuracy when the end-user is a trainee rather than a specialist2, 6. The deliberately compact architecture supports potential mobile and low-resource deployment. Clinically, such a tool could support overnight EEG monitoring by non-specialist staff, help medical students recognize epileptiform patterns, and give junior residents a quick reference during independent. The tool is not intended for unsupervised clinical deployment and must be used only under specialist supervision. Future priorities include patient-wise cross-validation; corrected annotation-based Siena labeling; multi-class seizure classification per ILAE 20173; temporal modeling with LSTM, GRU, or transformer layers19; ICA-based artefact rejection7; and prospective multi-center validation on a substantially larger cohort before any educational deployment15.
Disclosures
Nothing to disclose by any author(s). This research received no external funding. The eleven clinical EEG recordings from Aster Royal Al Raffah Hospital, Muscat, Oman were obtained under institutional governance and were fully de-identified prior to any computational processing under the supervision of the attending neurophysiologist (Clinical Supervisor, Dr. Achint Krishna). Data was used solely for non-commercial educational research validation and were not uploaded to any external server or shared repository. Public datasets (CHB-MIT and Siena) were accessed via PhysioNet under their data-use agreements permitting non-commercial research use. Ajitav Sahoo contributed in a personal capacity as technical advisor and not on behalf of his employer.
References
[1] Niedermeyer, E., & da Silva, F. L. (Eds.). (2004). Electroencephalography: Basic Principles, Clinical Applications, and Related Fields (5th ed.). Lippincott Williams & Wilkins.
[2] Sanei, S., & Chambers, J. A. (2007). EEG Signal Processing. Wiley.
[3] Fisher, R. S., et al. (2017). Operational classification of seizure types by the International League Against Epilepsy. Epilepsia, 58(4), 522–530.
[4] Scheuer, M. L., et al. (2017). Seizure detection: Interreader agreement and detection algorithm assessments using a large dataset. J. Clin. Neurophysiol., 34(5), 426–432.
[5] LeCun, Y., et al. (2015). Deep learning. Nature, 521(7553), 436–444.
[6] Cooley, J. W., & Tukey, J. W. (1965). An algorithm for the machine calculation of complex Fourier series. Math. Comput., 19(90), 297–301.
[7] Subasi, A., & Gursoy, M. I. (2010). EEG signal classification using PCA, ICA, LDA and support vector machines. Expert Syst. Appl., 37(12), 8659–8666.
[8] Baumgartner, C., & Koren, J. P. (2018). Seizure detection using scalp EEG. Epilepsia, 59(S1), 14–22.
[9] Kiral-Kornek, I., et al. (2018). Epileptic seizure prediction using big data and deep learning: Toward a mobile system. EBioMedicine, 27, 103–111.
[10] Guo, L., et al. (2010). Automatic epileptic seizure detection in EEGs based on line length feature and artificial neural networks. J. Neurosci. Methods, 191(1), 101–109.
[11] Schirrmeister, R. T., et al. (2017). Deep learning with convolutional neural networks for EEG decoding and visualization. Hum. Brain Mapp., 38(11), 5391–5420.
[12] Lawhern, V. J., et al. (2018). EEGNet: A compact convolutional neural network for EEG-based brain–computer interfaces. J. Neural Eng., 15(5), 056013.
[13] Roy, Y., et al. (2019). Deep learning-based electroencephalography analysis: A systematic review. J. Neural Eng., 16(5), 051001.
[14] Acharya, U. R., et al. (2018). Deep convolutional neural network for the automated detection and diagnosis of seizure using EEG signals. Comput. Biol. Med., 100, 270–278.
[15] Faust, O., et al. (2018). Deep learning for healthcare applications based on physiological signals: A review. Comput. Methods Programs Biomed., 161, 1–13.
[16] Shoeb, A. H. (2009). Application of machine learning to epileptic seizure onset detection and treatment (Doctoral dissertation). MIT.
[17] Zhang, Y., et al. (2020). Epilepsy seizure prediction on EEG using common spatial pattern and convolutional neural network. IEEE J. Biomed. Health Inform., 24(2), 465–474.
[18] Wang, X., et al. (2019). Detection analysis of epileptic EEG using a novel random forest model combined with grid search optimization. Front. Hum. Neurosci., 13, 52.
[19] Lee, D., et al. (2024). A ResNet-LSTM hybrid model for predicting epileptic seizures using a pretrained model with supervised contrastive learning. Sci. Rep., 14, 1319.
[20] Asif, U., et al. (2021). SeizureNet: Multi-spectral deep feature learning for seizure type classification. In *MLCN/RNO-AI 2020, LNCS 12449* (pp. 77–87). Springer.
[21] Zhou, M., et al. (2018). Epileptic seizure detection based on EEG signals and CNN. Front. Neuroinform., 12, 95.
[22] Ru, Y., et al. (2024). Combining data augmentation and deep learning for improved epilepsy detection. Front. Neurol., 15, 1378076.
[23] Craik, A., et al. (2019). Deep learning for electroencephalogram (EEG) classification tasks: a review. J. Neural Eng., 16(3), 031001.
[24] Guttag, J. (2010). CHB-MIT Scalp EEG Database (version 1.0.0). PhysioNet. https://doi.org/10.13026/C2K01R
[25] Detti, P. (2020). Siena Scalp EEG Database (version 1.0.0). PhysioNet. https://doi.org/10.13026/5d4a-j060
[26] Detti, P., et al. (2020). EEG synchronization analysis for seizure prediction: A study on data of noninvasive recordings. Processes, 8(7), 846.