Epilepsy affects an estimated 4 to 10 active cases per 1,000 people worldwide, with a disproportionately higher incidence in low- and middle-income countries, and it is characterized by recurrent seizures caused by sudden, abnormal surges of electrical activity in the brain that produce hypersynchronous neuronal firing. Because seizures can manifest across motor, cognitive, emotional, autonomic, and electroencephalographic domains, accurate detection is essential for timely intervention and improved patient outcomes. Electroencephalography remains the gold standard for non-invasive seizure detection, but visual inspection of EEG recordings is time-consuming, labor-intensive, and requires considerable expertise, and the low signal-to-noise ratio of many seizure events combined with the inherently non-stationary nature of EEG signals often complicates accurate identification of ictal activity. Deep learning has emerged as a powerful tool capable of extracting complex information from raw EEG data, and the new study builds on that promise while deliberately avoiding many of the methodological shortcuts that have inflated performance claims in the field.

The researchers, led by Raha Pedram and Hamid Abbasi alongside Alistair J. Gunn, Joanne O. Davidson, and Laura Bennet, introduced a compact temporal convolutional network, or TCN, comprising just 709,697 trainable parameters, roughly 0.71 million, which corresponds to a model size of approximately 2.71 megabytes. Unlike many conventional approaches that rely on extensive preprocessing or handcrafted feature engineering, this architecture processes multi-channel EEG directly without bandpass filtering, z-score normalization, notch filtering, or denoising, preserving the natural signal variability and characteristics of the recordings. The network employs depthwise-separable residual blocks with exponentially increasing dilation factors of 1, 2, 4, 8, and 16, allowing it to capture both short-term and long-range temporal dependencies across channels and time without recurrent structures. Each residual block also incorporates a squeeze-and-excitation module that adaptively recalibrates latent feature maps, while residual connections support gradient stability and enhance convergence, and spatial dropout mitigates co-adaptation of latent feature maps. The team selected the Swish activation function over the more common ReLU because Swish helps maintain smoother gradient propagation and preserves small negative activations that may carry diagnostic information in EEG signals.

To evaluate the model under the most challenging conditions, the researchers restricted analysis to the minimum common 18-channel EEG configuration consistently available across all subject sessions, extracted balanced 8-second windows with 50 percent overlap yielding 2,458 ictal and 2,458 interictal segments, and performed leave-one-recording-case-out cross-validation, or LORCO-CV, in which one CHB-MIT recording case was held out for testing while data from the remaining cases were used for training in each of 24 folds. Under this rigorous case-level scheme, the model achieved an average accuracy of 82.3 percent with a standard deviation of 16.5, precision of 83.5 percent, sensitivity of 80.1 percent, F1-score of 80.1 percent, and an area under the receiver operating characteristic curve of 0.866. These numbers are notably lower than the headline figures reported in many prior CHB-MIT studies, and the researchers argue that this is precisely the point: many earlier studies excluded difficult recording cases, often those corresponding to cases 12 through 16, or applied heavy filtering that can artificially inflate performance metrics without adequately addressing the generalizability required for real-world clinical deployment.

When the same model was evaluated under a complementary case-grouped 5-fold cross-validation, in which all windows from a given recording case were kept together with respect to the outer train-test split, performance rose dramatically to an average accuracy of 98.2 percent with a standard deviation of only 0.8, sensitivity of 98.3 percent, and an AUC of 0.999. A separate seizure-event-grouped 5-fold sensitivity analysis, designed to prevent windows originating from the same seizure event from being split across training, validation, and test sets, retained high sensitivity of 95.8 percent with an AUC of 0.952, indicating that the model’s discriminative capacity was not an artifact of within-event window dependence. An ablation analysis confirmed that both the squeeze-and-excitation modules and the dilated convolutions contributed meaningfully to the final architecture, with removing either component leading to modest but consistent reductions in accuracy, and a Wav2Vec2 Transformer baseline included as an internal comparator achieved lower accuracy of 76.9 percent under the same experimental setting.

The most striking finding emerged from the case-level stratification of results. The 24 recording cases descriptively grouped into three response sets: 15 good responders with accuracy above 80 percent, 5 moderate responders including cases chb04, chb06, chb08, chb15, and chb24 with accuracy between 65 and 80 percent, and 4 poor responders including chb05, chb12, chb13, and chb14 with accuracy below 65 percent. Two cases, chb13 and chb14, showed below-chance AUC values, indicating unreliable discriminative ranking for these subject sessions rather than consistent robustness across all recording cases. In a post hoc exploratory restricted-cohort analysis using only the 15 good responders, the model achieved 91.5 percent accuracy and 0.986 AUC, though the researchers emphasize that this analysis is descriptive rather than confirmatory and should not be treated as independent validation.

To understand why performance varied so substantially across cases, the team conducted a systematic spectral analysis using Welch’s method to estimate power spectral density on a per-channel basis, integrating the results within six canonical frequency bands: delta from 0.5 to 4 Hz, theta from 4 to 8 Hz, alpha from 8 to 13 Hz, beta from 13 to 30 Hz, gamma from 30 to 60 Hz, and high-gamma from 60 to 128 Hz. They prespecified three low-frequency spectral markers, the ictal-interictal delta difference, the ictal-interictal theta difference, and a normalized combined delta-theta index, and assessed their association with case-level AUC using Spearman’s rank correlation with bootstrap confidence intervals and Benjamini-Hochberg false discovery rate correction. The strongest association was observed for the theta-band difference, with a Spearman’s rho of 0.5779 and an FDR-adjusted p-value of 0.0083, followed by the normalized delta-theta index at rho of 0.5486 and the delta-band difference at rho of 0.4838. These associations remained statistically significant in participant-aware leave-one-session sensitivity analyses, indicating that the spectral findings were not driven by the repeated longitudinal observations represented by cases chb01 and chb21, which the dataset documentation indicates were recorded approximately 1.5 years apart from the same female participant.

The spectral analyses revealed that seizures were most reliably distinguished by concurrent increases in delta and theta power, with well-classified recording cases exhibiting robust low-frequency elevation in which delta rose more than five-fold and theta three- to four-fold relative to baseline. Conversely, poor responders demonstrated atypical profiles: some cases showed elevated baseline delta during non-seizure states that inflated false positives by mimicking ictal activity, while seizures in cases chb12 through chb15 lacked clear low-frequency enhancement, contributing to high false-negative rates. Band-power-by-channel heatmaps further suggested the presence of high-amplitude noise in certain recording cases, including abnormal gamma bursts in chb13 and low-frequency noise during interictal stages in chb04, which adversely influenced model performance. Morphological comparisons of misclassified segments highlighted interictal epochs with ictal-like waveforms that reduced class separability, emphasizing that robust seizure detection should include both spectral context and waveform morphology rather than relying on low-frequency power shifts in isolation.

An intriguing cohort-wide observation was a consistent elevation of delta and theta power in frontal channels, including FP1-F3, FP1-F7, FP2-F4, and FP2-F8, for both seizure and non-seizure events. The researchers note that this frontal low-frequency dominance may reflect a possible frontal low-frequency pattern or propagation of ictal activity to frontal sites in some cases, or alternatively persistent frontal confounds such as ocular potentials, sleep-related slow waves, or systematic recording factors including electrode contact and bedside movement. Because this frontal slow-power baseline appears in interictal data as well, it reduces ictal-interictal separability and can bias classifiers toward false positives unless explicitly accounted for. The team cautions that the CHB-MIT dataset does not provide dedicated electrooculography channels, sleep-stage annotations, movement recordings, expert artifact labels, or sufficient seizure-localization metadata to separate neurophysiological activity from recording-related confounds, so these patterns should be interpreted as signal-level observations rather than definitive evidence of a seizure-specific physiological marker.

Under a separate EDF-file-grouped 5-fold continuous recording-level evaluation designed to assess performance under naturally imbalanced EEG conditions, the model achieved event sensitivity of 91.0 percent with a standard deviation of 10.3, a mean detection latency of 15.9 seconds, window-level specificity of 92.5 percent, and a false-alarm rate of 3.21 per hour with an alarm-level positive predictive value of only 3.4 percent. The researchers applied a three-consecutive-window temporal-persistence criterion to reduce isolated positive classifications before alarm generation and a 120-second alarm-merging interval to avoid counting temporally fragmented detections as multiple alarm events, with all post-processing rules fixed before evaluation of held-out test recordings. These results indicate that the model detected most seizure events under continuous recording-level evaluation, but the relatively high false-alarm rate and low alarm-level PPV show that false positives remain a major limitation for deployment-oriented use, and the authors state plainly that the model is not yet suitable for clinical deployment without further false-alarm suppression.

The study’s authors argue that their work moves beyond the pursuit of classification accuracy alone by examining why performance varies across subjects and recording sessions, providing signal-level explanations for variability that has been consistently observed but rarely systematically investigated in the CHB-MIT literature. Cases identified as difficult in their analysis, such as chb05 and chb13, have also been reported as outliers in prior seizure-detection and seizure-forecasting studies, suggesting that these cases potentially reflect genuine recording-session-level signal complexity rather than isolated model failure. The researchers recommend that future work should standardize channel-subset protocols, develop channel-wise imputation or augmentation strategies, integrate artifact-aware preprocessing and training, employ case-specific normalization, and explore hybrid architectures that fuse temporal and spectral representations, potentially including Transformer-based modules augmented with frequency-domain features. They also note that uneven channel coverage, with recordings for chb13, chb16, and chb19 limited to 18 channels, can reduce spatial sampling compared with cases using the full electrode set and can further constrain topographic and spectral analysis.

The broader significance of the study lies in its demonstration that pediatric EEG seizure detection is influenced not only by model architecture but also by patient-specific and case-specific signal variability, and that performance estimates based only on cohort-level or window-level validation may obscure clinically important heterogeneity. By linking model outcomes with spectral characteristics across individuals, the researchers provide deeper insight into seizure pattern heterogeneity within the CHB-MIT dataset and address the often-overlooked question of why performance varies substantially across cases. Their compact architecture, comprising roughly 0.71 million trainable parameters, supports the use of the term lightweight, though the authors avoid claiming parameter-level superiority over all prior approaches since many related CHB-MIT studies do not report model size. The work was supported by the Health Research Council of New Zealand and conducted on the New Zealand eScience Infrastructure high-performance computing cluster using a single NVIDIA A100 GPU, with each training epoch requiring approximately three seconds to complete. The findings highlight the need for future seizure-detection systems to combine compact deep learning models with case-specific spectral and morphological interpretation while prioritizing false-alarm reduction before clinical translation, and they suggest that the delta-theta spectral signature identified in this study may serve as a useful marker for anticipating which patients’ recordings will be most amenable to automated detection.

Subject of Research: Pediatric EEG seizure detection using a lightweight temporal convolutional network: recording-case-level spectral insights from the CHB-MIT dataset

Article Title: Pediatric EEG seizure detection using a lightweight temporal convolutional network: recording-case-level spectral insights from the CHB-MIT dataset

Article References: Pedram, R., Gunn, A. J., Davidson, J. O., Bennet, L., & Abbasi, H. (2026). Pediatric EEG seizure detection using a lightweight temporal convolutional network: recording-case-level spectral insights from the CHB-MIT dataset. Results in Engineering, 32, Article 113259. https://doi.org/10.1016/j.rineng.2026.113259

Image Credits: AI Generated

DOI: Not provided

Keywords: Pediatric, seizure, detection, lightweight, temporal, convolutional, network, recording-case-level, spectral, insights, CHB-MIT, dataset