Seizures in newborns can be difficult to recognize even when a baby is being monitored continuously in an intensive-care unit. Their electrical signatures may be subtle, brief, or obscured by noise, while the sheer volume of electroencephalography (EEG) data places heavy demands on clinical specialists. A new computational study describes a hybrid artificial-intelligence system designed to identify the preictal state—the period preceding a seizure—from neonatal EEG while also showing which signals influenced its decisions. The model combines self-supervised contrastive learning, a neuromorphic spiking neural network, and five explainable-AI methods. Tested on recordings from 79 term neonates in the Helsinki University Hospital Neonatal EEG Seizure Dataset, the system achieved 90.39 percent accuracy, 90.02 percent recall for preictal segments, and an area under the receiver-operating-characteristic curve of 0.910. The researchers present the approach as a possible foundation for an early-warning tool that could operate on compact hardware in neonatal intensive-care units. It is not, however, a clinically validated diagnostic system: the evaluation was retrospective and based on a single dataset.
The clinical problem is consequential because delayed recognition of neonatal seizures can allow repeated abnormal electrical activity to continue before treatment begins. Newborn EEG is especially challenging to interpret: normal activity changes with developmental state, artifacts can resemble neurological events, and seizures may have limited visible clinical expression. The study notes that expert readers can miss roughly one in four events under standard monitoring conditions, consistent with the broader difficulty of visual interpretation reported in neonatal care. The researchers therefore focused not simply on detecting an ongoing seizure, but on classifying EEG segments as preictal or interictal, meaning sufficiently distant from a seizure to represent a non-seizure baseline. They defined preictal data as the four-minute interval before seizure onset and interictal data as periods more than five minutes from any seizure onset or offset. A 60-second guard interval and all ictal segments were excluded, preventing the two labels from overlapping. This produced a strongly imbalanced learning problem: interictal segments outnumbered preictal segments by approximately 6.48 to one.
The dataset contained about 5,800 hours of continuous, multichannel EEG from 79 term infants and 456 annotated seizure events. Signals were recorded through a 21-channel International 10–20 montage at 256 hertz, providing coverage across frontopolar, frontal, central, temporal, parietal, and occipital regions, along with auxiliary ECG and respiration channels. The researchers divided the recordings into overlapping 10-second epochs, generating 75,488 usable segments after preprocessing. Each channel was normalized separately within each recording to reduce differences in scale and signal drift, and flatline clips were set to zero. Crucially, the split was performed by patient rather than by individual epoch. Fifty-five infants were assigned to training, 12 to validation, and 12 to testing, so neighboring windows from the same recording could not appear in different partitions. The held-out test set contained 10,889 segments, including 9,376 interictal and 1,513 preictal examples. The authors also report a five-fold patient-level cross-validation analysis intended to test whether results depended too heavily on one division of the cohort.
The first stage of the model addresses a central limitation in medical AI: labeled seizure examples are scarce, while unlabeled monitoring data are abundant. Inspired by the SimCLR framework, the researchers used self-supervised contrastive pretraining primarily on interictal EEG. For each segment, the training process created two altered views and taught an encoder to produce similar representations for the paired versions while separating representations from other examples. The alterations were designed to mimic conditions encountered in clinical recordings, including Gaussian noise, temporal shifts, random channel dropout, pointwise masking, and amplitude scaling. A one-dimensional residual convolutional encoder transformed the 21-channel signals into a lower-dimensional representation. Its projection head produced a normalized 64-dimensional contrastive embedding. In this setting, the system did not need seizure labels to learn general features of neonatal EEG. According to the study, these pretrained representations improved downstream F1 scores by 8 to 12 percent compared with the relevant non-pretrained configurations, while the contrastive training loss fell below 0.1.
The second stage combines the learned representation with conventional signal-processing information before passing it to a spiking classifier. The pretrained module supplied 192 features: a 128-dimensional encoder output and a 64-dimensional contrastive projection. The researchers also calculated power spectral density with Welch’s method across five frequency bands—delta, theta, alpha, beta, and gamma—for each of the 21 electrodes. These 105 spectral measurements were compressed to 32 features, producing a 224-dimensional input. The classifier, called an attention-enhanced spiking neural network, used fully connected layers with batch normalization and dropout, followed by leaky integrate-and-fire neurons. These units accumulate input in a membrane-potential state, gradually lose that potential through leakage, and emit a binary spike when a threshold is reached. The network simulated this process over 50 timesteps, allowing it to represent temporal evolution rather than treating each input as a static vector. A surrogate gradient enabled backpropagation through the otherwise discontinuous spike-generation function, and average spike rates were used to produce probabilities for the preictal and interictal classes.
The architecture was trained with focal loss, which gives extra emphasis to difficult examples and the under-represented preictal class without discarding data through resampling. The model contained approximately one million parameters and exhibited reported spike sparsity of 15 to 20 percent, features the researchers associate with potential low-power, edge-device deployment. On the held-out test data, it identified 1,362 of 1,513 preictal segments, corresponding to the reported 90.02 percent recall, while missing 151. Its precision was 60.35 percent, yielding an F1 score of 72.25 percent; the macro-F1 score was 83.22 percent and the weighted F1 score was 91.14 percent. The confusion matrix included 8,481 true-negative classifications and 895 false positives. The precision-recall analysis produced an average precision of 0.76. These figures illustrate the trade-off at the heart of an early-warning system: prioritizing sensitivity can produce more alarms, some of which may not correspond to a genuinely approaching seizure. The authors describe the high recall as clinically attractive but acknowledge that false alarms could contribute to alarm fatigue.
Interpretability was built into the analysis rather than treated as an afterthought. The researchers applied Integrated Gradients, SHAP, LIME, saliency gradients, and attention profiling to preictal examples, then mapped the resulting attributions to the standard electrode layout. These methods answer related but different questions: which features change a prediction, which contribute globally, which matter for an individual example, where the output is most sensitive, and how the model’s internal weighting is distributed. Across the analysis, temporal regions accounted for approximately 40 percent of the reported contribution, central regions 25 percent, frontal regions 20 percent, parietal regions 10 percent, and occipital regions 5 percent. SHAP identified the T3 temporal-left channel, F8 frontal-right channel, and P4 parietal-right channel among the leading contributors. Temporal channels such as T3 and T4 and the central midline site Cz repeatedly ranked highly across attribution methods and sampled windows. The researchers say this pattern is qualitatively consistent with established descriptions of temporal and central-temporal involvement in neonatal seizure activity, but they emphasize that the explanations have not undergone formal validation by expert neurophysiologists.
The results suggest that combining representation learning, spectral information, and event-driven temporal modeling may help address the particular constraints of neonatal EEG, but substantial barriers remain before clinical use. The study was conducted offline on a single publicly available dataset, and performance on recordings from other hospitals, equipment, populations, and clinical workflows remains unknown. Fixed preictal windows may not represent the same biological process for every infant, motivating future adaptive or personalized definitions. Continuous explainability analysis could also be computationally demanding, even if the underlying classifier is compact. The authors propose further work involving model compression, lighter interpretability methods, multimodal information, streaming evaluation, and clinician-in-the-loop assessment. Ethical safeguards, patient privacy, and direct clinical oversight would be essential in any deployment. For now, the system is best understood as a research prototype: a promising attempt to forecast neonatal seizure-related activity while exposing the EEG regions and features behind its predictions, rather than as a replacement for specialist monitoring or medical judgment.
Contrastive pretraining is particularly relevant to neonatal EEG because the model can learn recurring structure from recordings that lack event annotations. By bringing augmented views of the same signal closer in representation space, the encoder is encouraged to retain features that remain stable despite modest shifts, noise, amplitude changes, or missing channels. This may improve robustness to routine recording imperfections, although the value of any augmentation depends on whether it preserves clinically meaningful seizure-related information. An alteration that is harmless for baseline EEG could potentially obscure a transient abnormality.
The spiking component provides a different form of temporal representation from the preceding convolutional encoder. A leaky integrate-and-fire unit carries a decaying internal state, so inputs separated in time can influence one another without requiring every signal value to be processed identically. The reported sparsity indicates that many potential spike operations are absent, which could reduce energy use on suitable neuromorphic hardware. It does not by itself establish faster or more efficient clinical operation, however, because total system cost also includes signal conditioning, feature extraction, memory access, and explanation generation.
Performance should also be interpreted at the level of clinical episodes rather than only short EEG windows. A high segment-level recall can arise when several neighboring epochs from one evolving event are correctly classified, while false positives distributed across long recordings may still create a burdensome alarm rate. Prospective testing would therefore need episode-level sensitivity, false alarms per monitoring hour, warning time, calibration, and stability across infants. Attribution maps can help investigate such behavior, but agreement among explanation methods is not proof that the highlighted electrodes represent a causal seizure mechanism. Their main immediate value is supporting model auditing and clinician review.
Subject of Research: Interpretable AI for forecasting neonatal seizures from EEG recordings
Article Title: A hybrid spiking neural network with contrastive pretraining for interpretable seizure forecasting using explainable AI
Article References: Selvaraj, J., Krishna, R., Gupta, A., & Guruviah, V. (2026). A hybrid spiking neural network with contrastive pretraining for interpretable seizure forecasting using explainable AI. Discover Informatics, 1(1), Article 8. https://doi.org/10.1007/s44564-026-00010-5
Image Credits: AI Generated
DOI: 10.1007/s44564-026-00010-5
Keywords: neonatal seizures, electroencephalography, seizure forecasting, spiking neural networks, contrastive learning, explainable AI, neuromorphic computing, neonatal intensive care, hybrid, spiking, neural, network
Cite Scienmag News
APA MLA Chicago
Scienmag. (August 29, 2026). AI Model Forecasts Neonatal Seizures While Revealing Its EEG Reasoning. https://scienmag.com/ai-model-forecasts-neonatal-seizures-while-revealing-its-eeg-reasoning/
Scienmag. “AI Model Forecasts Neonatal Seizures While Revealing Its EEG Reasoning.” Scienmag, 29 August 2026, https://scienmag.com/ai-model-forecasts-neonatal-seizures-while-revealing-its-eeg-reasoning/. Accessed 29 August 2026.
Scienmag. “AI Model Forecasts Neonatal Seizures While Revealing Its EEG Reasoning.” Scienmag. August 29, 2026. https://scienmag.com/ai-model-forecasts-neonatal-seizures-while-revealing-its-eeg-reasoning/
Copy citation Download RIS
Tags: AI-based neonatal seizure forecastingartificial intelligence in neonatal carechallenges in neonatal EEG interpretationcontrastive learningcontrastive learning for seizure predictionearly warning systems for neonatal seizuresEEG data analysis in neonateselectroencephalographyexplainable AIexplainable AI for EEG analysisHybridmachine learning accuracy in neonatal EEGneonatal EEG seizure detectionneonatal intensive careneonatal intensive care unit seizure monitoringneonatal seizuresnetworkneuralneuromorphic computingneuromorphic spiking neural networkspreictal state prediction in newbornsseizure forecastingspikingspiking neural networks
