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

One of the pivotal advances in neonatal care revolves around the integration of sophisticated monitoring systems powered by real-time analytics and AI algorithms. These systems allow for continuous, non-invasive assessment of vital parameters such as oxygen saturation, heart rate variability, and cerebral oxygenation. Through machine learning models trained on vast datasets, clinicians can now predict impending complications with remarkable accuracy, enabling earlier and more precisely targeted interventions. This paradigm shift from reactive to proactive care is fundamentally transforming neonatal intensive care units (NICUs) worldwide.

Parallel to enhanced monitoring, the development of ultra-premature infant support technologies has marked a watershed moment in neonatal medicine. Devices such as artificial placenta systems and extracorporeal membrane oxygenation (ECMO) tailored for neonates offer the potential to bridge survival during critical periods of lung immaturity. Recent innovations include bioengineered membranes capable of mimicking placental gas exchange more efficiently while reducing the risks of thrombosis and infection. These technological marvels are critical in extending the viability window for extremely premature infants, those born at the cusp of viability.

Genomic medicine is another arena witnessing explosive growth with profound implications for neonatal care. Advancements in rapid whole-genome sequencing (WGS) have empowered clinicians to diagnose congenital anomalies and genetic disorders within hours after birth. This rapid diagnosis allows tailored therapeutic strategies that can significantly alter disease trajectories. The convergence of genomic data with electronic health records and AI-driven predictive tools is enabling personalized medicine approaches that consider an individual neonate’s unique genetic makeup, environmental exposures, and clinical status—a triumvirate critical to optimizing outcomes.

Nutrition science within neonatology has also experienced revolutionary progress. The understanding of human milk’s immunomodulatory and neurodevelopmental properties has catalyzed the development of enhanced breast milk fortifiers and bioengineered milk alternatives that closely approximate natural breast milk in composition and functionality. These advancements mitigate risks such as necrotizing enterocolitis and support neurocognitive development, particularly in preterm infants who are vulnerable to nutritional deficits. Moreover, precision nutrition strategies, informed by metabolic profiling, are increasingly being utilized to customize feeding regimens in NICUs.

Simultaneously, the field is witnessing a surge in telehealth applications tailored for neonatal populations, expanding access to expert care far beyond traditional hospital environments. Remote monitoring coupled with virtual consultations connects multidisciplinary teams to neonatal patients in underserved or remote areas, ensuring timely intervention and continuous care. This decentralization is enhancing equity in neonatal health while alleviating the burden on tertiary care centers. Furthermore, tele-education platforms are bolstering knowledge dissemination among healthcare professionals, rapidly translating emerging research into clinical practice.

Despite these extraordinary technological strides, neonatal care continues to confront significant challenges. Among the foremost is the ethical complexity arising from the balance between aggressive life-support measures and quality of life considerations, especially in the context of extreme prematurity and severe congenital conditions. Clinicians, families, and ethicists are engaged in nuanced discussions to establish guidelines that respect patient autonomy, parental rights, and inclusive decision-making processes amid the inherent uncertainty of neonatal prognoses.

Another critical area demanding attention is the management of long-term morbidities associated with neonatal interventions. While survival rates have improved markedly, many neonates face persistent risks for neurodevelopmental impairments, chronic lung disease, and vision or hearing deficits. Current research is intensively focused on elucidating the pathophysiological mechanisms underlying these sequelae and developing neuroprotective strategies, such as therapeutic hypothermia or anti-inflammatory treatments, to mitigate long-term disabilities. The field is progressively adopting a holistic viewpoint that extends beyond survival to encompass quality and functional outcomes across the lifespan.

Environmental and social determinants of neonatal health represent an emergent frontier in contemporary care models. Socioeconomic disparities, maternal health, prenatal exposures, and access to healthcare resources are increasingly recognized for their profound influence on neonatal outcomes. Efforts to integrate social prescribing, community-based interventions, and policy reforms into neonatal care pathways are underway to address these upstream factors comprehensively. This approach underscores the intersectionality of clinical care with public health and social justice imperatives.

Artificial intelligence, beyond monitoring applications, is shaping neonatal diagnostics through imaging and pattern recognition. Advanced computer vision algorithms now assist in interpreting cranial ultrasounds, MRI scans, and even subtle facial phenotypes linked with genetic syndromes. These tools dramatically reduce diagnostic delays and physician workload, especially in high-volume NICU settings. Additionally, AI-driven predictive models are being employed to optimize ventilator management and medication dosing, contributing to safer, more personalized therapeutic regimens.

The interplay between inflammation and immune modulation in neonates presents another fertile research domain. Innovations in immunotherapy and anti-inflammatory agents tailored for premature infants who exhibit distinct immune profiles are emerging. The nuanced understanding of neonatal immune ontogeny is vital to crafting interventions that minimize infection risks without impairing the developmental trajectories of immune tolerance or exacerbating inflammatory injury, such as bronchopulmonary dysplasia.

The mobilization of big data and multi-omics integration—combining genomics, proteomics, metabolomics, and microbiomics—heralds a new horizon in unraveling the complex biology of neonatal diseases. These integrative approaches facilitate the identification of novel biomarkers and therapeutic targets. Longitudinal cohort studies harnessing these data modalities are beginning to elucidate the early-life origins of chronic diseases, thus providing pivotal insights that can inform preventive and therapeutic strategies from birth.

Furthermore, neonatal pharmacology is undergoing transformation with the advent of model-informed precision dosing. Physiologically-based pharmacokinetic (PBPK) models tailored for neonates account for the unique and rapidly changing physiology in this population, guiding safe and effective drug use. These empirical tools are particularly crucial given the limited data from traditional clinical trials involving neonates and the ethical constraints surrounding experimental therapeutics in this vulnerable group.

Innovations in non-invasive ventilation strategies and respiratory support are providing improved bridging therapies for neonatal respiratory distress syndrome. High-flow nasal cannula systems, non-invasive positive pressure ventilation, and aerosolized surfactant delivery are evolving to reduce the need for intubation and mechanical ventilation, minimizing ventilator-associated complications. These advances contribute substantially to decreasing the incidence and severity of bronchopulmonary dysplasia and improving overall respiratory outcomes.

Lastly, fostering a family-centered care model is revolutionizing NICU environments with profound psychosocial benefits. Encouraging parental involvement in daily care and decisions, utilizing developmental care principles, and creating supportive environments not only improve neonatal outcomes but also mitigate parental stress and anxiety. This holistic care philosophy is being amplified through designing NICU spaces that facilitate bonding, breastfeeding, and parental presence, reflecting an integrative approach that values the family unit as central to neonatal success.

In sum, neonatal care in the twenty-first century is at an exhilarating crossroads of unprecedented innovation, profound challenges, and transformative potential. The confluence of technology, personalized medicine, ethical reflection, and collaborative care models is poised to continue reshaping the landscape of neonatal medicine. As healthcare systems adapt and evolve, the ultimate goal remains steadfast: nurturing the most vulnerable lives with precision, compassion, and visionary science.

Article References:
Çeri, A., Gültekin, N.D. & Keskin, D.M. Neonatal care in the twenty-first century: innovations and challenges. World J Pediatr 21, 644–651 (2025). https://doi.org/10.1007/s12519-025-00927-1

DOI: July 2025