Summary: Researchers have applied machine learning to detect hidden neurological warning signs in the brain’s resting electrical rhythms, eliminating the need to capture active seizures to support an epilepsy diagnosis. An advanced pattern-recognition algorithm can identify subtle electroencephalogram (EEG) abnormalities associated with a genetic form of epilepsy with high accuracy, opening a path toward earlier pediatric intervention and noninvasive precision care.
The computational framework creates a tailored “dictionary” of recurring waveform motifs in baseline EEG recordings to reveal underlying brain changes. This approach points to objective biomarkers that could reduce diagnostic delays and family anxiety while guiding targeted treatment strategies.
Key Facts
- Diagnostic limitations of short EEGs: Standard clinical EEG sessions typically record around 20 minutes of brain activity, which makes it unlikely to capture spontaneous seizures and challenging for clinicians to find subtle, non-obvious signs by eye.
- Waveform dictionary approach: Instead of searching for overt seizures, the algorithm analyzes baseline EEG as if it were an unknown language. It identifies frequently repeating electrical motifs, learns their contextual meaning, and highlights anomalies that can elude human reviewers.
- Testing without seizures: To evaluate the method, researchers analyzed multi-day recordings from more than 40 mice, some carrying epilepsy-linked variations in the TSC1 gene. The algorithm examined baseline segments containing no seizure activity.
- High accuracy in genetic detection: The machine-learning model distinguished between genetic backgrounds and detected the TSC1 mutation with high accuracy in two of three mouse strains, relying only on resting brain-wave patterns.
- Clinical translation underway: Supported by the Delaware Clinical and Translational Research ACCEL Program, the team is adapting the technique for shorter pediatric EEGs collected during epilepsy evaluations at Nemours Children’s Health.
- Reducing family stress: Objective early biomarkers could shorten the anxious wait many families endure while monitoring for seizure onset and reduce uncertainty about when or whether seizures will appear.
- Toward precision treatment and monitoring: The investigators envision brain-wave typing that helps clinicians interpret treatment effects more accurately and supports continuous monitoring via wearable EEGs for epilepsy and related neurodevelopmental conditions such as autism and ADHD.
Source: University of Delaware
Epilepsy can be difficult to diagnose because seizures often do not occur during routine EEG recordings. To address this gap, researchers at the University of Delaware and collaborators have developed an artificial intelligence method that identifies discreet signatures in baseline EEGs that indicate underlying neurological changes.
In a proof-of-concept study in mice, the team demonstrated that their approach can detect subtle differences in EEG activity linked to a genetic form of epilepsy, even when no seizures are visible. The results, published in the Journal of Neural Engineering, support moving the technique into clinical testing with pediatric EEGs from Nemours Children’s Health.
A dictionary of brain waves
Clinicians commonly use EEGs to support an epilepsy diagnosis, but routine clinical recordings provide only a brief snapshot of brain activity. When a seizure is not recorded, diagnosis depends on much subtler EEG clues that are difficult to spot visually. The University of Delaware team developed an algorithm that learns the brain’s background rhythms by identifying waveform motifs that repeat across recordings and mapping their contextual roles. The resulting “dictionary” of waveforms serves as a structured feature set the model uses to distinguish normal from atypical patterns.
“By letting the algorithm learn recurring waveform structures, we can detect micro-patterns in baseline EEG that are effectively invisible during manual review,” said Austin Brockmeier, assistant professor in electrical and computer engineering and computer and information sciences.
Controlled tests in a mouse model
The collaboration began when Brockmeier presented his computational neuroscience work and connected with Amanda Hernan, a researcher who studies how altered brain activity affects cognition and learning in children with epilepsy. They tested the method on EEGs from more than 40 mice, some engineered with disease-linked disruptions to the TSC1 gene and others without. Recordings spanned multiple days for each animal and the analysis focused exclusively on segments with no seizure events.
Even with only baseline activity, the algorithm identified differences across strains and detected the TSC1 mutation with strong accuracy in two of the three genetic backgrounds tested. The results indicate that resting EEGs can carry measurable, interpretable signals of neurological genotype even in the absence of overt seizures.
“These findings suggest that waveform patterns in resting EEG hold useful biomarkers of neurological differences,” Hernan said.
Moving toward clinical application
With support from the Delaware Clinical and Translational Research ACCEL Program, the researchers are now applying their method to pediatric EEGs collected at Nemours Children’s Health. Pediatric clinical recordings are typically much shorter than controlled multi-day animal datasets and encompass a broader range of epilepsy types, but the team remains optimistic that the algorithm can extract early tracking biomarkers from these shorter sessions.
Early detection of electrophysiological indicators could enable faster intervention and reduce the prolonged uncertainty families face while waiting for visible seizures. Hernan notes that the psychological toll of anticipating seizures is substantial: “Seizures follow natural cycles, and without objective markers families can be trapped in a persistent state of anxiety.”
Improved brain-wave pattern recognition can also refine treatment decisions. If a clinician begins a new medication during a natural lull in seizure activity, they may overestimate its effectiveness. Objective EEG biomarkers would provide context about where a patient sits in their seizure cycle and help avoid misattributing natural variation to therapeutic benefit.
Looking ahead, the investigators expect this technology to support continuous monitoring with wearable EEGs and to be extendable to other neurological and neurodevelopmental conditions where altered baseline rhythms carry diagnostic or prognostic information.
“This work advances precision medicine for neurological disorders,” Brockmeier said. “Brain-wave typing could guide more personalized interventions and better long-term tracking for at-risk patients.”
Key Questions Answered:
A: The algorithm learns the brain’s characteristic background rhythms by building a customized dictionary of frequently occurring waveform motifs. Those waveform counts and patterns serve as features that reveal micro-patterns and genetic anomalies in ordinary baseline EEG that are not visible to the naked eye.
A: Pediatric EEGs provide much shorter recording windows and encompass a wider variety of epilepsy presentations than controlled animal studies. Despite these challenges, researchers are adapting the method to extract meaningful biomarkers from clinical-length recordings.
A: By mapping objective brain-wave patterns, the system can indicate whether changes in seizure frequency reflect natural cycles rather than a drug response, helping clinicians evaluate a treatment’s true effect.
Editorial Notes:
- This article was edited by a Neuroscience News editor.
- Journal paper reviewed in full.
- Additional context added by staff.
About this epilepsy and AI research news
Author: Marina Jones
Source: University of Delaware
Contact: Marina Jones – University of Delaware
Image: Image credited to Neuroscience News
Original Research: Closed access. “Interpretable EEG biomarkers for neurological disease models in mice using bag-of-waves classifiers” by Maria Isabel Cano Achuri, Montana Kay Lara, Khalil Abed Rabbo, Benjamin T. Wilson, Austin Meek, J. Matthew Mahoney, Amanda E. Hernan, and Austin J. Brockmeier. Journal of Neural Engineering. DOI: 10.1088/1741-2552/ae4d8c
Abstract
Interpretable EEG biomarkers for neurological disease models in mice using bag-of-waves classifiers
Objective.
Electroencephalograms (EEGs) record the brain’s electrical activity over time. Distinct waveform patterns—both rhythmic oscillations and transient spikes—can serve as phenotypic biomarkers that reflect genotype-specific neural activity. Identifying such markers is especially valuable when seizures are not observed, which is common in clinical evaluations and in animal models that display subtle neurological changes without overt epilepsy. This study explores whether genotypes can be predicted from long-term EEGs recorded from freely moving mice representing six groups defined by the presence or absence of a TSC1 gene knockout across three inbred strains with different genetic backgrounds.
Approach.
The team developed a machine-learning method that predicts individual genotypes from occurrence counts of short EEG waveform segments. A dictionary of representative waveforms is optimized for each genotype, and the vectors of waveform occurrence counts are used as features in logistic regression models to predict genotype.
Main results.
Using two-fold cross-validation of dictionary learning and leave-one-individual-out genotype prediction, the researchers found that pooling waveform counts over multi-hour segments enabled reliable strain prediction with 70% accuracy (95% CI 62–78) versus a chance rate of 38%. For two of the three strains (DBA2 and C57B6), classifiers detected the TSC1 knockout genotype with accuracies of 86% (95% CI 70–101) and 67% (95% CI 55–79), respectively, even though no overt seizures were observed. A state-of-the-art time-series classifier (Hydra) achieved higher strain classification (98%) and comparable genotype detection for those strains (86% and 71%) but lacks interpretability.
Significance.
These methodologies demonstrate that EEG waveform motifs can serve as interpretable phenotypes and that a bag-of-waves representation is a promising feature set for identifying epilepsy-related genotypes, supporting further development toward clinical applications.