The investigators screened LKB1-pathway genes using the public GEO dataset GSE143272 and then evaluated the leading candidate in a pentylenetetrazol-induced status epilepticus model in rats. SIK1 expression was significantly upregulated in the hippocampus of the status epilepticus animals, although the study did not establish that this increase represented greater kinase activity. The experimental work used AAV-mediated gene knockdown and overexpression, lentiviral transfection of primary hippocampal neurons, electroencephalography, whole-cell patch-clamp recording and Western blot analysis. This combination allowed the researchers to connect the computational screening with seizure-related measurements, molecular changes and cellular electrophysiology, while remaining within the scope of an induced animal model and laboratory cultures.

Manipulating SIK1 in the rat model produced opposing effects. AAV-mediated SIK1 knockdown prolonged seizure latency and reduced seizure duration, whereas SIK1 overexpression exacerbated seizure severity. The results therefore support a functional role for SIK1 in the experimental phenotype: lowering its expression was associated with less severe seizure outcomes, while raising it was associated with more severe outcomes. The findings do not by themselves establish that SIK1 causes human epilepsy or that targeting it would change seizure outcomes in people. They instead identify a candidate regulatory mechanism that merits further investigation in relevant experimental systems.

Protein-protein-interaction network analysis and mediation analysis identified NR4A2 as a downstream effector of SIK1. The molecular results were inverse: SIK1 knockdown increased NR4A2 expression, while SIK1 overexpression decreased it. Functional rescue experiments strengthened this relationship. Co-overexpression of SIK1 and NR4A2 reversed the pro-epileptic effect produced by SIK1 overexpression alone, showing that increased NR4A2 could counteract the SIK1-associated phenotype under the tested conditions. Together, the data support a model in which SIK1 negatively regulates NR4A2 and in which reduced NR4A2 expression contributes to the increased neuronal excitability associated with SIK1. The study does not show that the entire pathway is a universal driver of epilepsy; rather, it characterizes the axis in the datasets, rat model and neuron cultures examined.

Whole-cell patch-clamp recordings provided a cellular correlate for the seizure and molecular findings. In magnesium-free-treated primary hippocampal neurons, SIK1 knockdown or NR4A2 overexpression restored action-potential amplitude to near-normal levels. The authors interpreted these observations as evidence that the SIK1–NR4A2 axis promotes neuronal hyperexcitability through SIK1-mediated negative regulation of NR4A2. Conversely, the rescue and electrophysiological results indicate that increasing NR4A2 can offset effects attributed to SIK1 in this experimental system. The work suggests that the SIK1–NR4A2 axis may be a target for further epilepsy research, but the evidence presented is preclinical: it comes from a chemically induced rat model and cultured neurons and does not demonstrate a treatment effect in patients.

Subject of Research: SIK1–NR4A2 regulation of seizure severity and neuronal excitability in a rat status-epilepticus model

Article Title: The SIK1-NR4A2 axis: a core driver of epilepsy identified by integrated machine learning and experimental validation

Article References: Liu, X., Zheng, S., Yang, K., Sun, L., Su, L., Zhong, S., Liang, H., & Luo, X. (2026). The SIK1-NR4A2 axis: a core driver of epilepsy identified by integrated machine learning and experimental validation. Frontiers in Cell and Developmental Biology, 14, Article 1887268. https://doi.org/10.3389/fcell.2026.1887268

Image Credits: AI Generated

DOI: 10.3389/fcell.2026.1887268

Keywords: epilepsy, SIK1, NR4A2, LKB1 signaling, status epilepticus, neuronal excitability, machine learning, hippocampus, rat model, primary neurons, SIK1-NR4A2, axis

The finding offers a cell-level clue about age-related changes in the ability to solve new problems. Fluid intelligence was assessed as part of a broader series of cognitive tests administered before surgery, covering reasoning, memory, attention, language and fine motor coordination. After surgery, researchers used microscopy to identify and count immature astroglia in selected areas of the hippocampus, a brain region involved in learning and memory. Older patients generally had more of the abnormal cells, while greater cell numbers were also associated with lower fluid-intelligence scores. The investigators described the work as one of the first studies to associate astroglial abundance with human cognition.

Astroglia have traditionally been viewed primarily as support cells that help maintain a healthy environment around neurons and support their communication. Their role in cognition has received less attention than that of neurons and has often been investigated through animal models or indirect measures such as gene activity. The new research differs by examining human tissue directly and quantifying the presence of immature astroglia. “We found that older patients tended to have more abnormal immature astroglia—and that higher numbers of these cells were also associated with poorer performance on tests of fluid intelligence,” said first author Aswathy Ammothumkandy, PhD, of the USC Stem Cell lab.

The research was made possible by the unusual availability of tissue removed during surgery to treat severe, drug-resistant seizures caused by mesial temporal lobe epilepsy. The operation yielded hippocampal tissue from 29 patients, allowing the team to pair cellular observations with cognitive testing conducted beforehand. This design provided a rare opportunity to study individual cells in a human brain and compare them with measured reasoning performance. The approach also illustrates why the findings should be interpreted as an association: the small, selected surgical population had epilepsy, and the investigators did not show that removing or changing the cells would alter cognition.

Non-invasive tools commonly used to study brain function, including electroencephalography and magnetic resonance imaging, generally do not resolve individual cells well enough for this type of cell-count analysis. Microscopic examination can reveal cellular features in detail, but tissue from living people is rarely available for research outside specialized clinical circumstances. Here, the integration of clinical care, cognitive assessment, surgery and laboratory analysis gave researchers access to matched biological and behavioral information. The participants’ epilepsy and the surgical context remain central limitations on how broadly the results can be generalized, even though the study was supported in part by the National Institutes of Health and published in Frontiers in Cellular Neuroscience.

Mesial temporal lobe epilepsy can involve seizures and cognitive decline, making the population clinically relevant but also distinctive. The researchers said the findings could have implications beyond epilepsy, and they emphasized that aging affects everyone. “Aging is something we’re all familiar with and all impacted by—it’s a matter of when, not if,” said co-corresponding author Michael Bonaguidi, PhD, who leads the USC Stem Cell lab. “What this study provides is a new therapeutic target that could someday help us slow down or potentially even reverse age-related declines in our ability to solve new problems.” The statement identifies a proposed research direction rather than a demonstrated treatment.

No drug, procedure or clinical intervention was tested in the study, and the researchers emphasized that further work is needed to determine whether abnormal immature astroglia contribute directly to cognitive decline. The data also do not establish that the cells are a cause of the age-related pattern; they could be a marker, a consequence, or part of a broader process involving epilepsy, aging and brain function. The authors’ suggestion that astroglia could represent a target for preserving cognition should therefore be understood as a hypothesis for future investigation, not an available treatment or an established clinical strategy. The scope of the evidence is especially limited because all participants were people undergoing surgery for severe, drug-resistant seizures.

Ammothumkandy is now investigating molecular differences among immature astroglia that might reveal possible drug targets. By comparing cells from people with epilepsy with control samples, she aims to identify features that distinguish abnormal cells and explore ways to restore them to a healthier state. Another possible avenue for future investigation is neuromodulation, which uses electrical or magnetic stimulation to alter nervous-system activity. The broader goal is to understand whether healthy or abnormal astroglial states might become accessible therapeutic targets for cognitive conditions, including accelerated age-related decline in people with epilepsy. These proposed approaches are described in the release as directions for research rather than interventions validated in the present experiment.

Charles Liu, MD, PhD, director of the USC Neurorestoration Center and a co-author of the study, said the work could benefit patients while informing wider questions about brain function. “Understanding processes like cognitive decline, which can be accelerated in people with epilepsy, could help us improve these patients’ lives while also shedding light on broader questions about brain function that affect us all,” Liu said. The authors also proposed that astroglia could be a shared target for cognitive decline and seizure control, based partly on previous work by the team. The present experiment, however, directly evaluated the relationship between immature astroglial abundance, age and fluid intelligence; it did not test seizure outcomes or show that targeting astroglia would improve either cognition or seizures.

For now, the most defensible conclusion is that abnormal immature astroglia become more prevalent with age in the hippocampus of this surgical epilepsy cohort and are associated with weaker performance on tests of fluid intelligence. The study supplies direct human-tissue evidence that warrants further investigation into why these cells accumulate, what molecular changes define their abnormal state and whether they actively impair reasoning. It does not establish a general-population trend, a direct causal mechanism or an available therapy. Instead, it identifies a specific cellular feature for additional investigation, with the eventual aim of determining whether modifying astroglial biology could protect cognitive function or support better control of epilepsy. The findings are strongest as an association in the 29-person surgical sample and should not yet be read as proof that abnormal immature astroglia drive age-related cognitive decline in people without epilepsy.

Subject of Research: Associations between immature hippocampal astroglia, age and fluid intelligence in adults with epilepsy

Article Title: Immature brain support cells may be linked to human age-related cognitive decline

Article References: Immature brain support cells may be linked to human age-related cognitive decline. Original publication (Bibliographic details unavailable.)

Image Credits: AI Generated

DOI: Not provided

Keywords: astroglia, fluid intelligence, cognitive decline, hippocampus, epilepsy, brain tissue, aging, USC, human brain, cellular neuroscience, Immature, brain

Researchers compared two cohorts drawn from the hospital’s diabetes and immunology databases. The low-titer group comprised consecutive patients whose diabetes began between 2000 and 2005 and who had positive GAD65 antibody results below 2,000 IU/mL. The high-titer group included patients whose laboratory results reached or exceeded 2,000 IU/mL between 2013 and 2024. The analysts used the Mann–Whitney U test, chi-square or Fisher’s exact tests, and multivariable logistic regression to compare clinical and immunological characteristics.

Among the 150 high-titer patients, 32 (21.3%) had a neurological syndrome classified as GAD65 antibody-associated disease. Epilepsy was the most frequent presentation, accounting for 13 of the 32 cases (40.6%). Other classifications included stiff-person syndrome in 10 patients, overlapping neurological syndromes in six, cerebellar ataxia in two and ophthalmoparesis in one. One patient with a syndrome resembling progressive encephalomyelitis with rigidity and myoclonus also had antibodies against the glycine receptor, CV2 and Zic4, as well as a malignant thymoma; the investigators considered the glycine-receptor antibodies the most likely explanation and did not classify the case as GAD65-associated disease.

Diabetes was present in 128 high-titer patients (85.3%), including 126 with diabetes associated with latent autoimmune diabetes in adults and two with immune checkpoint inhibitor-induced diabetes. Seventeen of the 128 patients with diabetes (13.3%) had neurological GAD65 disease. Conversely, 112 of the 118 high-titer patients without that neurological involvement had diabetes (94.1%). Diabetes occurred in 9 of 13 patients with epilepsy (69.2%), four of 10 with stiff-person syndrome (40.0%) and two of six with overlap syndromes (33.3%), but the differences between syndromes were not statistically significant.

Neurological disease began at a younger mean age among high-titer patients with diabetes: 43.8 ± 13.5 years, compared with 55.4 ± 14.1 years among those without diabetes (p = 0.03). Patients with neurological involvement also had higher GAD65 antibody titers than those without it (Mann–Whitney U test, p = 0.013). Both groups had a median titer of 2,000 IU/mL, but the distribution among affected patients extended to markedly higher values, with a third quartile of 41,357 IU/mL rather than 2,000 IU/mL. The observed ranges were 2,000–7,320,000 IU/mL in patients with neurological disease and 2,000–1,580,000 IU/mL in those without it.

Clinical findings did not differ significantly by sex: neurological disease affected 13 of 60 men (21.7%) and 19 of 90 women (21.1%). In the multivariable analysis, male sex was associated with lower odds of being in the high-titer group rather than the low-titer group (odds ratio 0.363, 95% confidence interval 0.218–0.606; p < 0.001), while age was not independently associated with titer group. Diabetes, neurological disease and the number of additional autoimmune diseases could not be estimated in that model because of complete separation, with zero or near-zero frequencies in the low-titer cohort.

Cerebrospinal fluid samples were available for only 11 high-titer patients, including 10 with neurological disease. GAD65 antibodies were detected in cerebrospinal fluid in eight patients by both assays used, and two additional cases were positive by one method and confirmed at the reference laboratory. Oligoclonal bands were present in six of seven patients for whom those data were available, and these six also showed intrathecal antibody synthesis. Because cerebrospinal fluid testing was limited and incomplete, the findings do not establish its diagnostic performance across the entire high-titer cohort.

The investigators noted that earlier series had reported neurological disease in 82% of high-titer patients and in 94% of those with titers above 10,000 IU/mL. They suggested that the lower 21.3% observed here may reflect a broader, less selected population drawn from a laboratory database rather than patients already referred to neurological or endocrine clinics. They also found patients with epilepsy and titers below 10,000 IU/mL, indicating that a higher cutoff may improve specificity at the expense of sensitivity.

Based on these observational data, the authors recommend serial serum GAD65 antibody testing for patients with diabetes and high titers, together with prompt neurological referral when symptoms arise. They emphasized that the recommendations rest on a retrospective dataset with a limited number of neurological events. Other limitations included incomplete variables, nonuniform laboratory methods, cerebrospinal fluid testing only in symptomatic patients, and the fact that not every high-titer patient was evaluated by a neurologist trained in epilepsy or movement disorders. Prospective studies with standardized assessments were therefore needed to confirm the observed associations and identify more reliable predictors of neurological involvement.

Subject of Research: Clinical and neurological characteristics associated with high and low serum GAD65 antibody titers in retrospective patient cohorts

Article Title: Clinical characteristics associated with GAD65 antibodies: a comparative study between high and low GAD65-ab cohorts

Article References: Lombardo-Del Toro, P., Jaraba, S., Hernández-Pérez, G., Nedkova, V., Casasnovas, C., Jauma-Classen, S., Gonzalez-Mera, L., Gómez-Llopico, R., García-Serrano, L., Sala-Padró, J., Morandeira, F., & Falip, M. (2026). Clinical characteristics associated with GAD65 antibodies: a comparative study between high and low GAD65-ab cohorts. Frontiers in Neurology, 17, Article 1976777. https://doi.org/10.3389/fneur.2026.1976777

Image Credits: AI Generated

DOI: 10.3389/fneur.2026.1976777

Keywords: GAD65 antibodies, neurological autoimmunity, type 1 diabetes, stiff-person syndrome, epilepsy, cerebellar ataxia, autoimmune diabetes, retrospective cohort, antibody titers, Clinical, characteristics, associated

The abstract reports that treatment with adrenocorticotropic hormone, valproic acid, and oral prednisone was followed by seizure control and improvement in electroencephalography. The girl later died from an unclear cause. These are treatment-associated observations in one patient; the abstract does not report medication doses, duration of response, specific electroencephalographic findings, or evidence establishing which treatment produced the observed improvements. It also does not specify the cause of death.

Pathogenic variants in FRRS1L have been associated with developmental and epileptic encephalopathy 37 (DEE37; OMIM 616981). The abstract states that affected patients typically exhibit global developmental delay, epileptic encephalopathy, hyperkinetic movement disorder with choreoathetosis, spasticity, and rigidity, while noting that rare variants have been reported. The investigators also summarized previously reported FRRS1L variants and compared phenotypes among affected patients.

Within the scope of this single-patient case report, the findings document the girl’s variant, associated clinical features, treatment-associated seizure and electroencephalographic improvement, and unclear cause of death. The authors state that the case expands the genetic and phenotypic spectra reported in DEE37. However, the available evidence comes from one patient and cannot establish the variant’s effects across a wider population, separate the contributions of the three therapies, or determine the cause of death. The article was published as an open-access case report in BMC Pediatrics on 9 September 2026.

Subject of Research: A case report of a novel homozygous FRRS1L variant in a Chinese girl with developmental delay and epilepsy

Article Title: Identification of a novel homozygous FRRS1L variant in a Chinese girl with development delay and epilepsy: a case report

Article References: Liu, L., Song, X., Li, B., Jia, S., Wang, D., Qian, X., & Wang, Z. (2026). Identification of a novel homozygous FRRS1L variant in a Chinese girl with development delay and epilepsy: a case report. BMC Pediatrics. https://doi.org/10.1186/s12887-026-07671-z

Image Credits: AI Generated

DOI: 10.1186/s12887-026-07671-z

Keywords: FRRS1L, developmental delay, epilepsy, DEE37, whole-exome sequencing, genetic variant, case report, movement disorder, choreoathetosis, Identification, novel, homozygous

The study, published as a brief research report in Frontiers in Cellular Neuroscience, combined preoperative neuropsychological testing with analysis of surgically removed hippocampal tissue. Mesial temporal lobe epilepsy is the most common form of adult epilepsy, and people whose seizures are not controlled with medication may undergo removal of the epileptic hippocampus. That clinical setting creates a rare opportunity to connect cognitive performance measured before surgery with cellular abnormalities in the same individuals’ brain tissue. The investigators used it to test whether an unusual astroglial population previously identified in this patient group was associated with particular cognitive deficits.

Epilepsy is associated with accelerated brain aging and progressive cognitive impairment, with cognitive decline especially pronounced in older patients. The researchers examined whether the Dcx-positive, star-shaped cells in epileptic hippocampus tracked specific domains of cognition rather than serving as a nonspecific indicator of poor overall brain health. The targets included fluid intelligence and reasoning, visuospatial reconstruction, visuospatial memory, verbal learning and other functions assessed in the preoperative evaluation. The approach was made possible because these patients often have detailed cognitive testing and, when surgery is required, provide the hippocampal tissue needed for direct cellular examination.

The cognitive cohort included 40 Spanish-speaking adults with pharmacoresistant mesial temporal lobe epilepsy. Their average age was 39.88 years, their average disease duration was 26.74 years, and they had completed an average of 7.9 years of education. The researchers retrospectively analyzed the Neuropsychological Screening Battery for Hispanics together with a grooved-pegboard fine-motor test. Raw scores were converted to education-adjusted z-scores using education-based normative data, while the pegboard measure was normalized for age and education. The 15 tests or domains covered intelligence, attention, processing speed, language, visuospatial ability, visual and verbal memory, executive functioning and fine-motor dexterity.

Among the 15 measures, the clearest age-associated declines occurred in fluid intelligence and reasoning, visuospatial reconstruction and visuospatial memory. Verbal delayed recall, visual attention, processing speed and divided attention also showed significant correlations with age in the initial analysis, but those findings did not withstand correction for multiple comparisons using the Benjamini–Hochberg false-discovery-rate procedure. Regression models then examined age alongside disease duration. The combined model was significant for fluid intelligence, although neither variable reached conventional significance independently. Age had the stronger inverse association and lower p-value, accounting for 8.3% of residual variance after adjustment for disease duration, compared with 2.7% for disease duration. Neither age nor disease duration independently predicted the two visuospatial outcomes after adjustment.

For the cellular analysis, investigators examined 29 surgically resected hippocampi. Seventeen cases came from the team’s earlier work and 12 additional cases were processed through the same staining and quantification pipeline. Dcx-positive cells with a star-like shape were counted in three regions of the dentate gyrus: the granule cell layer, the hilus and the molecular layer. Prior work by the group established that all of these cells also expressed the astroglial markers GFAP and S100β, while they lacked co-expression with markers of granule neurons, mossy cells, interneurons, microglia and PSA-NCAM-positive immature neurons. The cells were therefore treated as a practical histological index of pathological immature astroglia, and the earlier study had not detected this population in control postmortem hippocampus.

Among the 29 patients, the number of Dcx-positive immature astroglial cells increased significantly with age in the molecular layer, with p = 0.0187. Counts in the granule cell layer and hilus showed no significant age-related change. A regression model supported this regional pattern: age at surgery was a positive predictor of molecular-layer cell counts and uniquely explained 17.3% of residual variance, while disease duration was not significant and accounted for 2.7%. Age and disease duration themselves were positively correlated, but disease duration was not associated with immature-astroglia counts in any of the three regions. The result therefore points to an age-associated accumulation of the abnormal cells in one hippocampal compartment rather than a uniform increase throughout the dentate gyrus.

The direct link with cognition was tested in 19 patients for whom both Spanish-language neuropsychological data and histology were available. Greater numbers of immature astroglia in the hilus and molecular layer were associated with lower Raven’s fluid-intelligence z-scores, with p values of 0.04 and 0.02, respectively. The association in the granule cell layer approached significance at p = 0.052. These correlations did not survive correction for multiple testing, and immature-astroglia counts were not significantly associated with visuospatial reconstruction, visuospatial memory or the other tested cognitive domains. The pattern was therefore specific to abstract, nonverbal reasoning in the available data, rather than a general relationship between cell abundance and every measure of cognition.

More complex regressions produced an important qualification. When age, disease duration and immature-astroglia counts were modeled together, the overall models were significant, but none of the individual predictors remained independently significant. Age contributed the largest unique share of residual variance in fluid intelligence, from 12.6% to 25.4%, while astroglial measures contributed 8.3% to 15.6% and disease duration only 3.3% to 4.1%. Models containing age plus astroglial counts retained a negative age–intelligence relationship, whereas the astroglial counts lost significance. In models with disease duration plus astroglial counts, both were negatively associated with fluid intelligence. The authors interpret these results as evidence that age and immature astroglia share some variance in relation to cognitive performance, while also cautioning against claiming an independent effect from such a small sample.

Several limitations constrain how far the findings can be generalized. The histological sample was small, the subset with paired cognitive data included only 19 patients, and the investigators did not conduct sex-based analyses because statistical power was insufficient. Seizure frequency and other potentially relevant clinical factors could not be examined because complete seizure records were unavailable. Surgery occurred within one year of cognitive testing for 10 of the 19 paired cases, but the remaining nine underwent surgery between one and five years afterward. In addition, Dcx positivity and star-like morphology provided a practical proxy rather than a comprehensive molecular definition. The authors recommend additional astroglial markers, single-cell sequencing and spatial transcriptomics for future work to characterize these cells and the signaling pathways associated with them.

Even with those caveats, the study offers a specific new lead for studying cognitive decline in epilepsy. It links an age-related cellular change in the hippocampus with fluid intelligence, a domain that also depends heavily on frontal brain networks. Whether hippocampal immature astroglia affect reasoning through disrupted hippocampal–frontal communication, or instead accumulate as a consequence of an already deteriorating brain, remains unresolved. The work also does not show that the same process occurs in healthy aging or in other neurological disorders. Future mechanistic and larger clinical studies will be needed to determine whether the cells are a cause, a consequence or part of a broader network of changes involving seizures, neuroinflammation, oxidative stress, blood–brain barrier leakage and neuronal hyperactivity. For now, the evidence identifies immature astroglia as a candidate cellular correlate—and possible research target—rather than a proven cause of declining intelligence in people with epilepsy.

Subject of Research: Association of hippocampal immature astroglia with age-related fluid-intelligence decline in people with mesial temporal lobe epilepsy

Article Title: An age-related decline in human fluid intelligence corresponds with immature astroglia prevalence

Article References: Ammothumkandy, A., Corona, L., Atai, N. A., Abedi, A., Jimenez, N., Armacost, M., D’Orazio, L. M., Zuverza-Chavarria, V., Cayce, A., Shariq, M., McGinn, R., Lee, D., Lee, B., Heck, C., Russin, J. J., Liu, C. Y., Smith, J. A. D., & Bonaguidi, M. A. (2026). An age-related decline in human fluid intelligence corresponds with immature astroglia prevalence. Frontiers in Cellular Neuroscience, 20, Article 1926804. https://doi.org/10.3389/fncel.2026.1926804

Image Credits: AI Generated

DOI: 10.3389/fncel.2026.1926804

Keywords: astroglia, fluid intelligence, mesial temporal lobe epilepsy, cognitive decline, aging, hippocampus, doublecortin, neuropsychology, dentate gyrus, epilepsy, age-related, decline

The framework, called CMS-NAS, combines three distinct families of neural networks within a search space specifically designed for seizure prediction: convolutional neural networks, the Mamba architecture and spiking neural networks. Rather than simply connecting the three types of networks, the researchers designed two hybrid building blocks, called ConvSpike and MambaSpike, which weave the complementary strengths of the components into single units. The system first converts multichannel EEG signals into time-frequency representations using a short-time Fourier transform. It then generates a population of candidate network architectures, each of which is trained and evaluated. Better-performing architectures are selected for crossover and mutation to produce new generations, an evolutionary loop that iteratively refines the candidate pool.

Crucially, the search simultaneously considers prediction performance and model size, allowing the algorithm to seek architectures that balance accuracy and efficiency. The resulting search space contains about 1.74 × 10¹⁷ possible architectures, a number so vast that manual exploration would be hopeless. By constraining the space around seizure prediction tasks and the hybrid building blocks, the team reduced the structural redundancy, search cost and computational burden typically associated with manually designed or conventional neural architecture search models. The approach reflects a broader trend in biomedical AI: instead of asking human engineers to guess the best network topology, the search algorithm itself explores the design landscape and evolves solutions tuned to the data.

To evaluate seizure prediction rigorously, the researchers considered several measures. AUC reflects the model’s overall ability to distinguish different EEG states, with higher values generally indicating better discrimination. Sensitivity measures the proportion of actual seizures that are successfully predicted, while the study’s false-positive rate measures the number of false alarm events per hour of interictal EEG. Higher sensitivity and lower FPR are therefore desirable. The researchers also calculated P values against a random prediction baseline, with P < 0.05 considered statistically significant. To reduce isolated false alarms, they used a k-of-n strategy, in which at least 20 of 30 consecutive predictions had to be positive to count as a positive event. The prediction horizon was set to 1 minute, giving patients and caregivers a short but actionable warning window.

The team tested CMS-NAS on two public datasets, CHB-MIT and AES, and on a private clinical dataset. They used both patient-specific and patient-independent evaluation. In the patient-specific setting, data from other seizure events of the same patient were used for training and validation, while one seizure event was held out for testing. In the patient-independent setting, all data from one patient were held out for testing while the remaining patients were used for training, providing a more demanding test of generalization to previously unseen patients. This dual evaluation matters because real-world clinical deployment would require models that work for new patients, not only for those whose data the system has already seen.

On CHB-MIT, CMS-NAS achieved an AUC of 0.930, sensitivity of 88.5% and an FPR of 0.108 false alarms per hour in patient-specific testing. Seventeen of 18 subjects met the study’s statistical significance criterion. In patient-independent testing, the model achieved an AUC of 0.672 and sensitivity of 60.7%, remaining ahead of the compared methods on these two core measures. On the AES dataset, CMS-NAS achieved an AUC of 0.878, sensitivity of 87.3% and an FPR of 0.107 in patient-specific testing, while its patient-independent sensitivity reached 61.3%, the highest among the compared methods. On the private clinical dataset, the model achieved an AUC of 0.923, sensitivity of 84.4% and an FPR of 0.138 in patient-specific testing, and in patient-independent testing it achieved an AUC of 0.662, sensitivity of 51.6% and an FPR of 0.253.

The researchers also evaluated deployment efficiency, and the numbers are striking for a field often dominated by heavyweight models. The architecture search took 11 hours, while the final model had only 0.034 million parameters, a model size of 0.131 MB, 0.293 million MACs, and an inference latency of 1.308 milliseconds. Such a footprint opens the door to running seizure prediction directly on lightweight wearable or bedside hardware, where memory, energy and latency budgets are tight. For patients, that could eventually translate into continuous, low-power monitoring outside the hospital rather than prediction confined to specialized clinical equipment.

Ablation experiments showed why the hybrid design works. CNN, Mamba and SNN were not simply stacked together but played complementary roles: the CNN extracted local time-frequency features, Mamba modeled long-range temporal dependencies, and the SNN used sparse spike-based processing to enhance transient abnormal signals. Removing any core component reduced overall prediction performance, confirming that the search algorithm had good reason to favor architectures containing all three. The result suggests that biologically inspired spiking computation, efficient sequence modeling and classical convolutional feature extraction each contribute something the others cannot replace when the task is spotting the fleeting electrical signatures of an impending seizure.

The researchers conclude that CMS-NAS can reduce some of the structural redundancy, search cost and computational burden associated with manually designed or conventional NAS-based seizure prediction models, while balancing prediction performance and lightweight deployment. However, the current framework is designed for fixed tasks and uses only single-modal EEG data. Future work will explore multimodal seizure prediction and architectures that can adapt to a wider range of tasks, further testing the framework’s generality and practical value. The paper, A Hybrid Convolutional, Mamba and Spiking Neural Network Architecture Search for Seizure Prediction, was published in the journal Cyborg and Bionic Systems on Sep 30, 2026, by authors including Bo Yu, Chang Li, Rencheng Song, Xiang Liu, Ruobing Qian and Xun Chen, with support from the National Key Research and Development Program of China, the National Natural Science Foundation of China and other Chinese funding agencies.

Subject of Research: Neural architecture search for EEG-based epileptic seizure prediction

Article Title: AI searches for a better, lightweight network to predict epileptic seizures

Article References: AI searches for a better, lightweight network to predict epileptic seizures. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: epilepsy, seizure prediction, EEG, neural architecture search, deep learning, spiking neural networks, Mamba, convolutional neural networks, CHB-MIT, lightweight AI, Cyborg and Bionic Systems, biomedical engineering

The case centers on an Iranian girl referred to the Genetic Research Center in Meybod, Yazd, with a constellation of symptoms that had resisted every conventional diagnostic approach. She presented with developmental delay, low muscle tone, and refractory epilepsy — seizures that persisted despite multiple anticonvulsant medications. For her family, the diagnostic odyssey had been long and exhausting, a familiar story for parents of children with rare genetic conditions who often spend years undergoing test after test without ever receiving a definitive answer.

To crack the case, the team led by Noorodin Karami of Shahid Sadoughi University of Medical Sciences turned to whole exome sequencing, a technique that reads the roughly one to two percent of the genome containing the instructions for building proteins. This targeted approach captures the most likely disease-causing variants while keeping the enormous data volume of whole genome sequencing manageable. The sequencing revealed something remarkable: a novel frameshift variant in the CDKL5 gene, never before documented in any patient anywhere in the world.

The variant, designated c.1037_1038delTCinsCAGATCTAA, p.Ile346Thrfs*14, is what geneticists call a frameshift mutation. Rather than simply swapping one DNA letter for another, it disrupts the reading frame of the gene — the triplet code that determines how the cellular machinery translates DNA into protein. Think of it as deleting a word in the middle of a sentence and inserting nine letters in its place: everything downstream becomes gibberish. In this case, the mutation introduces a premature stop codon just 14 amino acids after the disruption point, truncating the CDKL5 protein far shorter than its normal length and stripping away essential functional domains it needs to do its job in the brain.

CDKL5 stands for Cyclin-Dependent Kinase-Like 5, a serine-threonine kinase located on the X chromosome. The protein it encodes acts as a molecular switch, adding phosphate groups to other proteins to regulate their activity. Among its known targets is Cav2.3, a voltage-gated calcium channel that plays a crucial role in controlling neuronal excitability. When CDKL5 phosphorylates Cav2.3, it alters the channel’s inactivation kinetics — essentially changing how quickly the channel closes and resets. Disruption of this process can shift the delicate balance of electrical signaling in the brain, potentially lowering the threshold for seizures. The protein also contains nuclear localization signals and a nuclear exit signal, indicating it shuttles between the nucleus and cytoplasm, where it influences gene expression and other cellular processes critical for healthy brain development.

Because the girl is an only child with no family history of similar illness, the researchers performed Sanger sequencing — a gold-standard validation technique — on her parents to determine whether the variant was inherited or had arisen spontaneously. The results confirmed the mutation was de novo, meaning it appeared for the first time in the child and was absent from both parents. This pattern is common in CDKL5 deficiency disorder, where most affected individuals carry mutations that no one else in the family has. The finding also carries important implications for genetic counseling: since neither parent carries the variant, the recurrence risk for future pregnancies in this family is extremely low, though not zero due to the theoretical possibility of germline mosaicism.

The team classified the variant as pathogenic according to the American College of Medical Genetics and Genomics criteria, the international standard for evaluating disease-causing potential. Bioinformatic analyses further supported this classification, predicting that the truncation would trigger nonsense-mediated mRNA decay — a cellular quality control mechanism that destroys mRNA transcripts containing premature stop codons before they can be translated into protein. The consequence is a double hit: the truncated protein that is produced lacks essential domains, and much of the mRNA never even survives long enough to be translated. The net result is a profound loss of CDKL5 function in the developing brain.

On the treatment front, the case yielded a clinically significant observation. After trying various medication combinations, the team found that valproate combined with lamotrigine produced a considerable reduction in the number of seizures. Valproate is a broad-spectrum anticonvulsant that works through multiple mechanisms, including enhancing GABA signaling — the brain’s primary inhibitory neurotransmitter system — and blocking voltage-gated sodium channels. Lamotrigine, meanwhile, primarily blocks sodium channels and reduces glutamate release, the brain’s main excitatory neurotransmitter. The combination targets both sides of the excitation-inhibition balance that goes awry in epilepsy. While CDKL5 deficiency disorder has no cure and the seizures rarely disappear entirely, meaningful seizure reduction can dramatically improve quality of life for affected children and their families.

CDKL5 deficiency disorder, sometimes described as a cousin of Rett syndrome, affects an estimated one in 40,000 to 60,000 live births worldwide, though the true prevalence may be higher due to underdiagnosis. The condition typically manifests with early-onset epileptic seizures within the first months of life, followed by severe developmental delay, hypotonia, and a range of other neurological problems. Because the gene sits on the X chromosome, females are more commonly affected, though males with CDKL5 mutations can also develop the disorder. The clinical spectrum is broad, and genotype-phenotype studies have begun to map how specific mutation locations correlate with disease severity — information that helps clinicians anticipate complications and tailor management strategies.

The Iranian case expands the mutational spectrum of CDKL5 and reinforces a message that geneticists have been championing for years: sequencing-based genetic testing should be considered early in the diagnostic workup of children with refractory epilepsy and developmental delay. Early molecular diagnosis not only ends the diagnostic odyssey but also enables personalized therapeutic management, informed genetic counseling, and access to emerging targeted therapies now in development for CDKL5 deficiency disorder. As research into this condition accelerates globally — from zebrafish models revealing dysregulated gene expression to international consensus recommendations for clinical management — each newly identified variant adds another piece to the puzzle, bringing the field closer to understanding this rare and relentless disease.

Subject of Research: Identification and characterization of a novel de novo CDKL5 frameshift variant causing infantile-onset refractory epilepsy and developmental delay in an Iranian child

Article Title: A novel de novo variant of CDKL5 in an Iranian child with infantile-onset refractory epilepsy

Article References: Karami, N., Dehghani, M., Asadollahi, S., Ghasemi, N., Movahedinia, M., & Kalantar, S. M. (2026). A novel de novo variant of CDKL5 in an Iranian child with infantile-onset refractory epilepsy. Molecular Biology Reports, 53(1), Article 1729. https://doi.org/10.1007/s11033-026-12884-7

Image Credits: AI Generated

DOI: 10.1007/s11033-026-12884-7

Keywords: CDKL5, epilepsy, novel variant, whole exome sequencing, developmental delay, pediatric neurology, genetics, refractory seizures, novel, novo, variant, Iranian

The scientific case rests on a remarkable convergence of biological properties. Zebrafish embryos develop outside the mother, with most organs forming within 24 hours of fertilization, and remain optically transparent for the first few days of life. This allows researchers to watch organogenesis, tumor growth, immune cell migration, and cardiac function unfold in real time inside a living vertebrate. A single breeding pair produces hundreds of eggs per week, fish reach sexual maturity in roughly three months, and maintenance costs are a fraction of those for mammalian colonies. Crucially, around 80 percent of human disease-linked genes have zebrafish orthologs, and core molecular processes—signal transduction, endocrine regulation, immune responses, and xenobiotic metabolism—are highly conserved, including functional counterparts of key drug-metabolizing enzymes and transporters. Modern genome-editing tools such as CRISPR/Cas9, Cre-lox systems, and transgenesis have all been adapted for the fish, enabling precise disease modelling and high-throughput in vivo screening at a scale unattainable in rodents.

The ethical architecture of the recommendations is anchored in the 3Rs framework of Replacement, Reduction, and Refinement, now extended by proposals for a fourth R, Responsibility, and a fifth R, Rehoming. The Society specifically recommends using embryos and larvae up to the stage preceding independent feeding—commonly around five days post-fertilization—for in vivo experiments. This timing matters legally as much as ethically: EU Directive 2010/63/EU defines its scope of protection with reference to independently feeding larval forms, meaning zebrafish embryos and early larvae fall outside the Directive’s formal scope. Polish national legislation, through the Act of 15 January 2015 on the protection of animals used for scientific or educational purposes, mirrors this exclusion. In practical terms, experiments on zebrafish embryos and early larvae may constitute a genuine legal alternative to classical protected-stage experiments, contributing to the replacement or reduction of mammalian models where scientifically appropriate. When later developmental stages or adult fish are required, the Society insists on approval from the competent Local Ethics Committee for Animal Experiments.

Nowhere is the model’s power more vivid than in oncology. The Society recommends the human-zebrafish xenograft approach, in which human cancer cells—from established lines or directly from patients—are implanted into zebrafish embryos or larvae. Because the larval immune system is not yet fully functional during the first days of life, human cells engraft and proliferate without immunosuppression. The transparent body permits direct visualization of tumor development, angiogenesis, and metastasis at single-cell resolution, and unlike in mouse models, human cancer cells in zebrafish readily migrate and form metastases, allowing drug effects on this critical aspect of tumor biology to be assessed under true in vivo conditions. Xenograft experiments can be completed in less than a week, whereas analogous mouse studies typically require several weeks. This so-called Avatar strategy is already entering clinical contexts: in a translational study involving breast cancer patients, complete concordance was demonstrated between chemotherapy responses observed in the larval model and the actual clinical responses, confirming high predictive value for personalized medicine. The model has also delivered clinically validated drugs—larval zebrafish studies revealed thalidomide’s strong inhibitory effect on tumor angiogenesis through suppression of VEGF signaling, helping establish thalidomide and its derivatives as therapies for neovascularization-dependent cancers such as multiple myeloma.

In toxicology, the recommendations position zebrafish as a tiered pre-filter for chemical and pharmaceutical safety. The Fish Embryo Acute Toxicity Test, standardized under OECD TG 236, determines lethal concentration values over a 96-hour exposure window in embryos, providing early hazard identification without adult vertebrates. While LC50 endpoints are not directly equivalent to rodent LD50 values, they deliver comparable information on acute toxicity and are primarily used for screening and prioritization rather than direct regulatory substitution. Embryonic transparency enables direct observation of teratogenic and embryotoxic potential, revealing morphological abnormalities, oxidative stress, apoptosis, and endocrine disruption induced by pharmaceuticals or contaminants. The assay’s compatibility with multi-well plate formats, robotic microscopes, computerized readers, and artificial intelligence allows automated detection of embryo orientation and organ-specific analysis of the heart, brain, or vessels. Regulatory recognition is growing: the FET assay is incorporated into OECD guidelines as an accepted alternative for chemical safety assessment, the US Food and Drug Administration acknowledges zebrafish for investigative and early-screening preclinical procedures, and the European Medicines Agency has expressed strong interest in preliminary regulatory reviews. The Society cautions, however, that zebrafish should not be regarded as a universal replacement where regulatory frameworks demand mammalian data on pregnancy, lactation, multigenerational endpoints, or long-term systemic organ toxicity.

Neurology and neuropsychiatry showcase the model’s translational depth. Zebrafish possess a hypothalamic–pituitary–interrenal axis homologous to the mammalian stress axis, with conserved cortisol synthesis, alongside well-organized serotonergic, dopaminergic, glutamatergic, and GABAergic neurotransmitter systems. Epilepsy research has been transformed by two complementary approaches: seizures can be induced in larvae with pentylenetetrazole, and genetic models based on the scn1lab gene—which corresponds to the human SCN1A gene mutated in Dravet syndrome—faithfully recapitulate reduced GABAergic signaling, spontaneous seizures, and abnormal neuronal activity. High-throughput phenotypic screening in multi-well formats using automated locomotor tracking and local field potential recordings identified clemizole as a potent seizure suppressor in the scn1lab model, and the same pipeline has highlighted serotonergic modulators such as fenfluramine and unexpected compound classes including synthetic cannabinoids. Classical antidepressants like fluoxetine and citalopram induce behavioral effects in zebrafish consistent with their human actions, and standardized larval assays—open-field thigmotaxis, light–dark preference, and light–dark challenge tests—quantify anxiety-like behavior with high predictive value. Social behavior assays in older larvae and juveniles have even been applied to model neurodevelopmental conditions relevant to autism spectrum disorders.

Metabolic disease research exploits the model’s conserved endocrine machinery. Zebrafish retain homologous insulin, glucagon, nuclear receptors, and appetite-regulating pathways; more than 70 percent of carbohydrate and glucose metabolic genes are conserved across zebrafish, frog, chicken, mouse, and human, with 57 percent linked to human type 2 diabetes. Larval insulin resistance models generated through elevated glucose, dexamethasone, or exogenous insulin exposure display features of impaired glucose handling suitable for evaluating antidiabetic drug candidates. For type 1 diabetes, transgenic lines expressing nitroreductase under the insulin promoter enable inducible, highly specific beta cell destruction upon metronidazole exposure, producing reproducible insulin deficiency and hyperglycemia while allowing real-time visualization of pancreatic islet regeneration. Adult fish on high-fat or high-cholesterol diets develop obesity, dyslipidemia, and atherosclerosis-like vascular lipid accumulation with conserved lipoprotein metabolism and inflammatory pathways. Storage diseases are equally tractable: the zebrafish model of Fabry disease, lacking alpha-galactosidase A activity, revealed an independent mechanism of kidney damage driven by mitochondrial alterations and disturbed glycolysis and galactose metabolism, beyond the canonical lysosomal accumulation of globotriaosylceramide.

Immunology and cardiology round out the model’s portfolio with distinctive technical advantages. During the first week of life, zebrafish larvae possess only a functional innate immune system, with adaptive immunity developing weeks later—a temporal separation that permits selective investigation of neutrophil and macrophage responses without adaptive interference. Transgenic reporter lines such as Tg(mpx:GFP) for neutrophils and Tg(mpeg1:mCherry) for macrophages enable direct visualization of leukocyte migration and host–pathogen interactions in living organisms. Infection of larvae with Mycobacterium marinum, closely related to Mycobacterium tuberculosis, recapitulates the hallmark organized granulomas of human tuberculosis and has enabled identification of virulence factors, host defense mechanisms, and high-throughput screening for novel antitubercular drugs, while reporter lines tracking autophagy marker LC3 recruitment to bacterial vacuoles illuminate cell-autonomous immunity against Staphylococcus aureus and Salmonella Typhimurium. In cardiology, homologs of key cardiac ion channels including hERG make zebrafish respond to QT-prolonging drugs comparably to humans: a landmark study identified 22 of 23 clinically relevant proarrhythmic compounds as positive in the assay, a drug-sensitized screen uncovered 15 genes affecting cardiac repolarization including GINS3, and screening of 3,000 compounds in a doxorubicin-induced cardiomyopathy model yielded the cardioprotective molecules visnagin and diphenylurea, subsequently validated in mice. Adult zebrafish uniquely regenerate cardiac muscle after injury through cardiomyocyte dedifferentiation and proliferation—a capacity absent in mammals—making the fish an unmatched system for heart regeneration research.

The recommendations arrive against a sobering backdrop of preclinical failure. The overall preclinical-to-market attrition rate exceeds 92 percent, a figure unchanged for three decades. Mouse models produce positive preclinical results in 90 to 95 percent of studies, yet fewer than 8 to 10 percent translate to approved drugs, falling to 3.5 percent in oncology—a gap attributed less to mouse biology than to study design, publication bias, sex bias, lack of genetic diversity, and failure to reproduce the full complexity of human immunology and metabolism. Notably, zebrafish under 120 hours post-fertilization, which do not swim or feed independently, are considered New Alternative Methodologies under EU Directive 2010/63, and the FDA has developed a draft on NAM use in preclinical research paving the way for zebrafish in these studies. The Society frames the fish as the most productive early-stage preclinical pipeline: an intermediate platform positioned between simple in vitro systems and resource-intensive mammalian studies, used to prioritize compounds, identify early toxicity signals, and refine hypotheses before mammalian validation. Implementation, the authors stress, requires question-driven design, standardized protocols with defined developmental stages, exposure windows, and controls, robust biologically relevant endpoints, and predefined criteria for progression to mammalian models. With zebrafish now the second most frequently used animal model in EU research after mice, the Society calls on research institutions, pharmaceutical companies, and regulatory agencies to systematically incorporate these embryo and larval assays into efficacy and safety evaluation, grant applications, and research standards—an integration the authors argue will yield more humane, efficient, and innovative science, ultimately benefiting medical progress and public health.

Subject of Research: Use of the zebrafish (Danio rerio) model in biomedical research and its recommendations for replacing mammalian models

Article Title: Recommendations of the Polish Zebrafish Society on the use of the zebrafish (Danio rerio) model in biomedical research

Article References: Podlasz, P., Migocka-Patrzałek, M., Prajsnar, T. K., Sarosiak, A., & Tylzanowski, P. (2026). Recommendations of the Polish Zebrafish Society on the use of the zebrafish (Danio rerio) model in biomedical research. Acta Biochimica Polonica, 73, Article 16545. https://doi.org/10.3389/abp.2026.16545

Image Credits: AI Generated

DOI: 10.3389/abp.2026.16545

Keywords: zebrafish, Danio rerio, 3Rs principles, preclinical research, drug discovery, toxicology, oncology, xenograft, cardiotoxicity, epilepsy, diabetes, immunology

The appeal of EEG as a diagnostic window is obvious: it is non-invasive, inexpensive, and captures neuronal dynamics in real time with millisecond precision. But the raw signal itself is notoriously difficult to work with. Scalp EEG operates in the microvolt range, and its five canonical frequency bands—delta (0.5–4 Hz), theta (4–8 Hz), alpha (8–13 Hz), beta (13–30 Hz), and gamma (above 30 Hz)—are easily swamped by physiological contamination. Eye blinks generate voltage shifts up to 100 microvolts concentrated over frontal electrodes, while jaw and neck muscle activity inject broadband noise above 20 Hz. Left uncorrected, low-frequency ocular artifacts can masquerade as the slow delta rhythms associated with brain injury, and muscle tension can obscure the gamma-band degradation that signals early dementia.

Traditional linear filters often fail because aggressive noise removal can inadvertently suppress the very neural features that carry diagnostic weight. Modern pipelines therefore deploy hybrid signal-processing strategies: Independent Component Analysis decomposes multi-channel recordings into statistically independent components that can be selectively removed, while wavelet-based variants isolate ocular artifacts without erasing underlying frontal delta rhythms. More exotic decompositions, such as Variational Mode Decomposition and Fourier–Bessel series expansion empirical wavelet transforms, target complex broadband muscle artifacts. Principal Component Analysis complements these methods by compressing high-dimensional, redundant channel data and suppressing noise-dominated components before classification.

Deep learning has transformed this denoising landscape by learning nonlinear, data-driven mappings from noisy signals to clean targets without requiring reference channels. Convolutional neural networks treat multi-channel EEG as spatial-temporal grids, sliding learnable filters along the time axis to extract shift-invariant features like epileptic spikes or rhythmic oscillations. Their highly parallelizable structure makes them uniquely suited for low-latency bedside warning systems, though their localized receptive fields struggle to model long-range temporal context or spatial relationships between non-contiguous electrodes. Recurrent architectures, including long short-term memory networks and gated recurrent units, maintain hidden states that carry historical context across time steps, making them effective for isolating slow physiological drift—but their sequential processing prevents efficient GPU parallelization, and a single high-amplitude artifact can derail the hidden state trajectory, cascading distortion across subsequent samples.

Autoencoders take a different approach, forcing contaminated signals through a tight information bottleneck that suppresses redundant noise while retaining stable neural components. Denoising variants are trained by deliberately corrupting inputs and forcing reconstruction of clean originals, learning noise-invariant feature mappings. Sparse autoencoders impose penalties that drive most latent neurons toward zero, encouraging each to specialize in distinct features, while variational autoencoders model the latent space as a continuous probability distribution. The principal limitation is that these models are extremely data-specific: an autoencoder optimized on uniform resting-state data will severely distort signals from sleep or cognitive tasks, and the inherently lossy compression can mask subtle diagnostic markers.

Generative adversarial networks frame artifact removal as a competitive game between a generator that transforms contaminated inputs into clean estimates and a discriminator that evaluates whether outputs resemble pristine EEG. Advanced iterations include CycleGANs, which denoise unlabeled datasets where paired clean data is unavailable, and Wasserstein GANs, which stabilize training through distance metrics. But because the generator’s only goal is to fool the discriminator rather than follow biological rules, these models are prone to mode collapse and can fabricate waveforms that look real but are entirely artificial—a serious concern when true uncorrupted human brain signals are rarely available as ground-truth supervision in clinical settings.

Transformer-based models bypass recurrent loops entirely, using self-attention mechanisms to dynamically weight different temporal segments and capture long-range dependencies across channels and time. Specialized variants process window-based sequences, treat time-frequency spectrograms as image-like inputs, or target noise signatures directly in the frequency domain. Their computational workload, however, grows quadratically with signal length, demanding massive memory for high-density multi-channel data, and they require exceptionally large datasets to avoid simply memorizing training examples. Most recently, large language model-inspired architectures treat EEG sequences as structured temporal tokens, enabling self-supervised pretraining and zero-shot or few-shot learning across heterogeneous subjects. The review is careful to delineate clinical readiness: automated report generation from pre-extracted features is empirically demonstrated, but direct autonomous diagnosis from raw signals remains a prospective application lacking rigorous multi-center validation.

Benchmark results compiled in the review show deep architectures achieving 91.5 to 96.8 percent classification accuracy on datasets like CHB-MIT and TUH EEG, compared to 84.0 to 89.5 percent for traditional support vector machines or random forests on identical raw inputs. Hybrid CNN-LSTM models have reached 99.7 percent accuracy for distinguishing Parkinson’s patients from healthy controls, while deep ensemble learning achieved roughly 97.9 percent average accuracy for Alzheimer’s classification. Yet the authors caution that deep learning does not universally supplant conventional approaches: on small cohorts of fewer than 50 subjects, classical pipelines pairing wavelet transforms or Hjorth parameters with SVM classifiers frequently match or outperform shallow CNNs while offering superior interpretability and far lower computational overhead.

The most damaging problem the review identifies is not architectural but methodological: artificially inflated performance metrics caused by data leakage. When continuous EEG records are split randomly by time segments rather than keeping each patient’s data entirely separate, models memorize individual biometric signatures instead of learning genuine disease indicators—a failure that becomes apparent only when testing on new subjects, age groups, or hospital equipment. The authors argue the field must transition to rigid subject-independent testing, leave-one-subject-out cross-validation, and cross-dataset evaluation, reporting sensitivity, specificity, balanced accuracy, and F1-scores rather than accuracy alone. They also stress that denoising and diagnosis must be evaluated separately, since aggressive cleaning can erase tiny chaotic brain waves that look like noise to a reconstruction network but contain the hidden keys to accurate diagnosis.

Looking forward, the review identifies explainable AI workflows, hardware-efficient edge models for bedside deployment, and rigorous empirical validation of foundation models as the priorities that will determine whether deep learning-driven EEG analysis fulfills its clinical promise. With neurological disorders currently affecting nearly one billion people worldwide—epilepsy alone accounting for 50 million cases and dementia over 24 million—the stakes of getting this translation right could hardly be higher. The tools exist; the challenge now is proving, under the unforgiving conditions of real hospitals and real patients, that machines can read the brain’s electrical whispers reliably enough to be trusted with diagnosis.

Subject of Research: Deep learning-driven EEG analysis for neurological disease diagnosis

Article Title: Computational approaches to deep learning–driven EEG analysis for neurological disease diagnosis: a systematic review

Article References: H F, T., & Jacob, J. E. (2026). Computational approaches to deep learning–driven EEG analysis for neurological disease diagnosis: a systematic review. Frontiers in Neuroinformatics, 20, Article 1778980. https://doi.org/10.3389/fninf.2026.1778980

Image Credits: AI Generated

DOI: 10.3389/fninf.2026.1778980

Keywords: EEG, deep learning, neurological disease diagnosis, epilepsy, Alzheimer's disease, Parkinson's disease, convolutional neural networks, transformers, generative adversarial networks, autoencoders, data leakage, explainable AI

The technique at the center of the research is stereoencephalography-guided radiofrequency thermocoagulation, or SEEG-RFTC. In this procedure, depth electrodes implanted under robotic navigation serve double duty: they record seizures from inside the brain, and then they deliver focal radiofrequency energy through selected contacts, heating tissue to as much as 80 degrees Celsius and creating small thermal lesions roughly seven millimeters across. It converts a purely diagnostic implantation into a therapy without a single additional surgical trajectory — an appealing option for patients whose seizure networks overlap eloquent cortex or who decline open craniotomy.

But which patients actually benefit? The Nanjing team, led by Y.-q. Yang and colleagues, turned to the seizure-onset pattern — the specific electrographic sequence that appears at the very start of a seizure on SEEG. Among the patterns classified in prior work, one stands out: low-voltage fast activity, or LVFA, a burst of rapid, low-amplitude rhythms that many epileptologists regard as the fingerprint of the epileptogenic core itself. Pathophysiologically, LVFA is thought to reflect local inactivation of inhibitory interneurons and synchronized firing of pyramidal cells — a disinhibition mechanism that optogenetic experiments have directly implicated. Slow waves and electrodecremental patterns, by contrast, are widely considered propagation territories rather than the source.

The study’s design exploits a unique property of RFTC: because the thermal lesions are strictly confined to the electrode contacts that recorded the seizure onset, there is an intrinsic spatial co-localization between the recorded brain-wave pattern and the clinical response. Unlike open resection, where the extent of tissue removal confounds interpretation, RFTC provides a clean model to test whether LVFA carries independent prognostic significance. The researchers analyzed 25 patients who received SEEG-RFTC as their sole intervention between 2022 and 2024, all with at least 12 months of follow-up and a mean observation period of 31.6 months.

The results are remarkably clear-cut. Among the 13 patients whose seizures began with an LVFA-containing pattern, 12 — 92.3 percent — achieved complete seizure freedom, graded Engel class I. Among the 12 patients with LVFA-absent patterns, only 4 — 33.3 percent — did so, a difference the authors report as statistically significant. The stratification held even when the team separated patients by MRI status: in MRI-negative patients, LVFA-present cases achieved 100 percent seizure freedom versus 20 percent for LVFA-absent cases, while MRI-positive patients showed 85.7 versus 42.9 percent. Notably, MRI positivity alone did not predict outcome at all.

The MRI-negative finding may be the most clinically provocative. When no structural lesion is visible on 3-Tesla epilepsy-protocol MRI, LVFA frequently corresponds to occult focal cortical dysplasia — particularly bottom-of-sulcus dysplasia, a malformation tucked into sulcal depths that evades even modern imaging. The authors suggest that MRI-negative patients with LVFA onset, especially when PET hypometabolism concords with the SEEG-defined onset zone, may form an electro-metabolic double-positive subgroup with exceptionally favorable ablation outcomes. Such patients, they argue, should not be excluded from RFTC solely on the basis of MRI negativity.

Yet seizure freedom, as graded by the Engel classification, is an inherently subjective measure. It relies on patient-reported seizure counts and recall, which can be inaccurate for subtle or nocturnal events. More fundamentally, Engel class I does not guarantee that the epileptogenic network has been completely disintegrated — residual subclinical electrical activity may persist and later evolve into clinical seizures, particularly when medications are tapered. This uncertainty complicates decisions about antiseizure medication withdrawal and long-term prognosis, leaving patients and clinicians without an objective readout of whether the driver has been adequately destroyed.

Here the study introduces its second key contribution: postoperative scalp EEG as an objective, noninvasive, longitudinally trackable complement to Engel grading. At six months after RFTC, 20 of the 25 patients underwent 24-hour scalp EEG recordings using the international 10–20 system, independently reviewed by two clinical neurophysiologists blinded to the seizure-onset patterns and clinical outcomes. EEG normalization — the absence of interictal epileptiform discharges without focal slowing beyond the expected postoperative range — was significantly more frequent in LVFA-present patients, appearing in 7 of 8 normal-EEG cases versus 4 of 12 abnormal-EEG cases. Seven of the 8 patients with normalized EEG were seizure-free, a positive predictive value of 87.5 percent.

The dissociations the team observed are as informative as the concordances. One LVFA-absent patient with delta brush onset achieved EEG normalization yet remained Engel class III — proof that scalp EEG and seizure counts capture complementary dimensions of therapeutic success. Conversely, among seizure-free patients with persistent EEG abnormalities, only half belonged to the LVFA-present group, suggesting incomplete network disruption despite clinical remission. The authors propose that Engel class I with a normal EEG may represent a more complete therapeutic outcome than Engel class I with persistent discharges, and might guide safer medication withdrawal. Among non-LVFA patterns, delta activity, beta sharp activity, and electrodecrement yielded zero seizure freedom, consistent with their interpretation as propagation rather than core phenomena — though sharp activity, an exception, achieved 100 percent seizure freedom in three patients while still lacking EEG normalization, hinting at a potentially less durable mechanism.

The authors are careful to frame these conclusions as preliminary. The cohort is small, the design retrospective, five patients lacked postoperative EEG, and the 64 percent Engel I rate in the RFTC-only subgroup reflects natural selection — the 18 patients who proceeded to open resection did so precisely because RFTC failed, whereas some non-seizure-free patients declined further surgery or were satisfied with reduced seizure burden. Across all 43 patients who initially received RFTC, the overall seizure-free rate was 37.2 percent, consistent with the 23 to 50 percent reported in meta-analyses. Still, the proposed decision framework — favoring RFTC as potentially definitive therapy when SEEG reveals LVFA, and urging caution and early evaluation for second-stage resection when slow-wave or electrodecremental patterns dominate — offers a concrete, testable path forward. If larger prospective studies confirm that targeting LVFA reliably produces both seizure freedom and scalp EEG normalization, epilepsy surgery may gain something it has long lacked: an objective brain-wave readout of whether the seizure generator has truly been silenced.

Subject of Research: SEEG-guided radiofrequency thermocoagulation outcomes stratified by seizure-onset pattern, with postoperative scalp EEG as an objective complement to Engel classification

Article Title: Complementing Engel Classification: Postoperative Scalp EEG as an Objective Indicator in SEEG‐Guided Radiofrequency Thermocoagulation—A Seizure‐Onset Pattern‐Based Stratified Analysis—

Article References: Yang, Y.-Q., Zhang, R., Ma, H., Yang, L., Qu, X., Chen, Q., & Zhang, Y. (2026). Complementing Engel Classification: Postoperative Scalp EEG as an Objective Indicator in SEEG ‐Guided Radiofrequency Thermocoagulation—A Seizure‐Onset Pattern‐Based Stratified Analysis—. Annals of Clinical and Translational Neurology, Article acn3.70553. https://doi.org/10.1002/acn3.70553

Image Credits: AI Generated

DOI: 10.1002/acn3.70553

Keywords: epilepsy, SEEG, radiofrequency thermocoagulation, low-voltage fast activity, scalp EEG, Engel classification, seizure-onset pattern, drug-resistant epilepsy, focal cortical dysplasia, epilepsy surgery, Complementing, Engel