Dual-Domain Deep Learning Network Accurately Detects Diagnostic Epilepsy Artifacts in Brainwaves

A multidisciplinary research team has developed an advanced dual-branch neural network capable of automatically detecting and classifying complex electroencephalogram artifacts, distinguishing routine physiological interference from subtle seizure-related automatisms. Published in Acta Epileptologica, the investigation introduces the Local-Global Feature Fusion Network Based on Time-Domain and Time-Frequency Domain, designated LG-TDTFD-Net. Electroencephalography, particularly continuous video electroencephalography monitoring, remains the primary diagnostic instrument for evaluating epilepsy, a chronic neurological disorder characterized by recurrent disruptions in electrical brain activity that affects approximately 50 million individuals worldwide. Temporal lobe epilepsy represents the most prevalent classification of focal epilepsy identified in tertiary clinical centers, accounting for roughly 66 percent of focal diagnoses, and constitutes the primary form of epilepsy evaluated for surgical intervention.

Recording cerebral electrical potentials via scalp electrodes inherently yields low-amplitude signals that remain highly vulnerable to distortion from external and internal sources. During extended continuous monitoring, these non-cerebral electrical transients frequently mimic true paroxysmal cerebral discharges, introducing significant risks of diagnostic misinterpretation and clinical error. Historically, computational neuroscientists concentrated almost exclusively on artifact removal through linear regression models and blind source separation algorithms such as independent component analysis. However, in temporal lobe epilepsy, seizures originating within temporal structures frequently manifest with stereotyped motor automatisms. Distinguishing these ictal behaviors from routine, voluntary chewing or normal ocular movements poses an enduring challenge, especially when patients inadvertently shift into camera blind spots during continuous monitoring. Automating this differentiation alleviates extensive clinician workloads while protecting diagnostic accuracy against oversight during long-term diagnostic recordings.

Previous computational attempts to automate artifact detection have revealed substantial technical limitations across algorithmic generations. Traditional machine learning pipelines relied heavily on manual feature engineering paired with classical classifiers, such as support vector machines for binary eye movement discrimination, linear logistic regression for five-class artifact sorting, or k-nearest neighbor models operating on fixed sliding windows before recurrent filtering. These conventional pipelines depended on static feature selection protocols and parameter optimization that lacked adaptability across variable recording montages. Subsequent deep learning architectures utilized convolutional neural networks to extract temporal patterns through localized sliding kernels, yet the constrained receptive fields of convolutional operations prevented networks from capturing long-range temporal dependencies. Conversely, while recurrent architectures such as long short-term memory networks addressed sequential modeling, their strictly recurrent operations prohibited parallel processing during model training. More recently, attention-driven transformer architectures succeeded in mapping global sequence relationships, but reliance on pure self-attention mechanisms generated smoothing effects that obscured high-frequency local transients critical for distinguishing fine waveform morphologies.

To overcome these architectural trade-offs, the investigators formulated LG-TDTFD-Net around a parallel dual-branch framework that processes raw one-dimensional voltage traces alongside dynamic two-dimensional spectral representations. The model receives two simultaneous inputs from each signal window: the raw time-domain sequence and a computed time-frequency spectrogram. The time-domain feature extraction module employs a tiered one-dimensional convolutional structure wherein the initial convolutional layer incorporates a wide kernel designed to capture broad baseline trajectories across extended intervals. This is immediately succeeded by a second convolutional layer equipped with a smaller kernel calibrated to register rapid signal fluctuations and local waveform contours. The integration of contrasting temporal scales allows the convolutional stage to extract rich multi-scale representations of complex physiological wave patterns. To stabilize gradient propagation and inhibit overfitting during optimization, batch normalization layers follow each convolutional layer, paired with exponential linear unit activation functions and subsequent max-pooling stages that downsample temporal dimensions while preserving salient morphological transients.

Following initial one-dimensional convolution, the extracted local feature sequence passes into a six-layer transformer encoder to construct global representations across the entire temporal window. The encoder incorporates five parallel attention heads configured with an embedding dimension of 40, enabling simultaneous modeling of long-range and short-range contextual relationships within independent mathematical subspaces. The query, key, and value matrices compute normalized similarity scores via scaled dot-product operations, utilizing a square-root scaling factor based on the key vector dimensionality to stabilize numerical gradients across long sequences. Unlike traditional transformer implementations that feed attention outputs directly into classification layers, the LG-TDTFD-Net routes the transformer feature sequence through a subsequent one-dimensional convolutional operation, denoted CONV1Db. This secondary convolution operates directly on the global contextual representations, extracting deep local features that refine the smoothed attention states. This sequence synthesizes three distinct descriptive tiers: shallow local features from initial filtering, global dependencies from multi-head self-attention, and refined deep local patterns extracted from the synthesized sequence.

Operating in parallel with the time-domain pathway, the time-frequency domain feature extraction module maps the temporal evolution of spectral energy across individual signal epochs. The raw one-dimensional signal undergoes short-time Fourier transform processing using a 250-point Hamming window with a 200-point temporal overlap, establishing a balanced compromise between frequency resolution and computational overhead. The resulting two-dimensional spectrogram, displaying frequency distributions over time, is processed by a specialized two-dimensional convolutional neural network. The initial convolutional layer utilizes a directional kernel measuring three points along the frequency axis and one point along the time axis, isolating spectral features, while the second layer applies a kernel measuring one point along the frequency axis and three points along the time axis to isolate temporal shifts. After batch normalization, activation, and pooling, the flattened and transposed feature matrix enters a second transformer encoder to map global cross-frequency dynamics. A subsequent two-dimensional convolution extracts deep local spectral patterns, completing an identical three-tiered hierarchy of shallow local, global, and deep local representations within the frequency domain.

The integration of these diverse analytical pathways occurs inside a specialized feature fusion module designed to preserve complementarity while minimizing computational complexity. Within each individual domain, the concatenated vectors representing shallow local, global, and deep local features are integrated using a one-dimensional convolutional layer with a kernel size of one. This unit-width convolution allows dense cross-feature interaction and dimensional compression without disrupting temporal sequence ordering. The compressed time-domain representations and time-frequency representations are subsequently concatenated into a unified multidimensional feature matrix encapsulating temporal morphology alongside evolving spectral power. This master feature representation passes into a classification head comprising two fully connected dense layers that project the combined feature space into categorical output dimensions corresponding to target artifact classes. A final softmax activation function converts output logit vectors into normalized class confidence distributions, optimized via categorical cross-entropy loss and the Adam optimization algorithm using a learning rate of 0.0002, exponential decay parameters beta-one of 0.9 and beta-two of 0.999, and a batch size of 200 samples.

Rigorous validation required testing the architecture across standardized benchmarking databases and authentic clinical recordings. The public benchmark was derived from the Temple University Artifact Corpus, comprising 310 long-term clinical recordings totaling 99.98 hours, re-referenced to a standardized bipolar montage and sampled at 250 hertz across five established categories: eye movements, muscle activity, chewing, electrode artifacts, and shivering. Shorter intervals failed to encompass complete artifact waveforms, whereas extending duration to five seconds expanded parameter volume from 4.5 million to 23 million without yielding performance gains. To evaluate clinical generalization, researchers collected an independent clinical dataset at the Third People’s Hospital of Chengdu, comprising 23 continuous one-hour video electroencephalograms recorded from temporal lobe epilepsy patients across frontal and temporal derivations. These raw 256-hertz recordings were resampled to 250 hertz via anti-aliasing interpolation filters, bandpass filtered between 0.5 and 80 hertz, and notch-filtered at 50 hertz to eliminate power line contamination.

Extensive ablation experiments confirmed the statistical necessity of every architectural tier within the unified framework. Incorporating the initial transformer encoder into a baseline convolutional network elevated overall accuracy, recall, and F1-score by 6.7 percent, 6.7 percent, and 5.8 percent, respectively, verifying the importance of global dependency modeling. Introducing intermediate convolutional fusion generated subsequent gains of 3.0 percent in accuracy and recall, while extracting deep local features via secondary convolution supplied additional improvements of 3.0 percent in accuracy and recall alongside a 5.0 percent rise in F1-score. Incorporating the parallel time-frequency branch yielded final marginal enhancements of 3.1 percent in accuracy and recall. When tested against existing classification frameworks on the public benchmark, LG-TDTFD-Net achieved an overall accuracy of 94.1 percent, a recall of 0.941, and an F1-score of 0.943.

Evaluating clinical utility, the researchers deployed the network to differentiate routine physiological movements from pathological seizure automatisms on the clinical hospital dataset. In a challenging six-class diagnostic task distinguishing muscle interference, regular eye movement, standard chewing, electrode artifacts, epileptic rhythmic blinking, and ictal rhythmic chewing, LG-TDTFD-Net attained an average classification accuracy of 96.0 percent, a recall of 0.960, and an F1-score of 0.961. The authors noted that the investigation remains restricted to retrospective laboratory analyses, requiring prospective validation before real-time deployment. Future translational hurdles include developing vendor-agnostic software interfaces to bridge disparate electroencephalography recording hardware, accommodating closed proprietary commercial platforms, and validating performance across non-temporal epilepsies exhibiting distinct cortical discharge dynamics. Furthermore, because LG-TDTFD-Net functions as an automated detection and classification instrument rather than an artifact subtraction filter, downstream signal restoration remains necessary for pure waveform analysis. Nevertheless, by transforming deceptive motor artifacts into actionable diagnostic markers, the dual-domain framework establishes an objective foundation for future clinical decision support systems.

Subject of Research: Deep learning network for electroencephalogram artifact detection and epileptic automatism classification

Article Title: Automatic EEG artifact detection using a local-global feature fusion network in time and time-frequency domains

Article References: Yu, Q., Lin, C., Wang, W., Feng, Z., & Wen, M. (2026). Automatic EEG artifact detection using a local-global feature fusion network in time and time-frequency domains. Acta Epileptologica, 8(1), Article 7. https://doi.org/10.1186/s42494-026-00248-4

Image Credits: AI Generated

DOI: 10.1186/s42494-026-00248-4

Keywords: Electroencephalography, Temporal Lobe Epilepsy, Deep Learning, Transformer Networks, Convolutional Neural Networks, Artifact Detection, Automatisms, Time-Frequency Analysis, Automatic, artifact, detection, local-global