AI Reads Brain Waves to Detect Hidden Brain Injury in Critically Ill Children

Every day in pediatric intensive care units around the world, clinicians face an agonizing uncertainty: a child lies sedated and encephalopathic, and no one can say for certain whether the brain is being silently injured. Computed tomography and magnetic resonance imaging offer snapshots, but they are costly, sometimes impractical, and often slow to reveal damage in its earliest stages. Now a prospective study conducted at Children’s National Hospital suggests that the answers may already be flowing through the electrodes taped to a child’s scalp. By combining routine electroencephalography (EEG) features with basic clinical data and machine learning classifiers, a team of pediatric neurologists and biomedical engineers has shown that algorithms can predict the presence of acute cerebral injury in critically ill children with an area under the receiver operating characteristic curve of 0.95 and a test accuracy of 0.90.

The research, published in the journal Neurocritical Care, enrolled 201 children admitted to the pediatric intensive care unit (PICU) between July 2021 and January 2023. The investigators deliberately excluded patients with preexisting cerebral injury or known epilepsy, ensuring that the models were learning to detect new, acute damage rather than simply recognizing old scars. Of the children studied, 42 percent were female and the median age was 3.5 years, with an interquartile range of 1 to 11.6 years, reflecting the wide developmental spectrum that any pediatric prediction tool must span. Acute cerebral injury was ultimately detected in 51 percent of the cohort, making the classification problem nearly balanced and therefore clinically meaningful.

The technical core of the study lies in how the researchers translated the raw electrical symphony of the brain into numbers a machine could digest. Rather than feeding continuous EEG waveforms directly into deep neural networks, the team extracted background features from each recording, classifying them according to standardized critical care EEG terminology. The most common background pattern in the cohort was slow and disorganized activity, seen in 73 percent of patients. Epileptiform discharges appeared in 24 percent of recordings, and frank electrographic seizures in 13 percent. These features were joined by clinical variables including demographics, the reason for PICU admission, and neuroimaging findings, creating a compact but information-dense feature vector for each child.

With this feature set in hand, the researchers ran a bake-off of six classical machine learning approaches: K-nearest neighbor, logistic regression, support vector machines (SVM) with both linear and radial basis function (RBF) kernels, random forest, gradient boosting, and Adaboost classifiers. The team designed two separate prediction tasks. The first was a binary classification, asking simply whether acute brain injury was present or absent. The second was a more ambitious multiclass problem, distinguishing no injury from unilateral injury and from bilateral injury, the kind of lateralization information that can guide urgent imaging decisions and targeted therapy.

For the binary task, the support vector machine with an RBF kernel emerged as the clear winner, achieving an AUROC of 0.95 and a test accuracy of 0.90 for predicting cerebral injury. SVMs work by finding the optimal hyperplane that separates classes in a high-dimensional feature space; the RBF kernel allows the decision boundary to bend into nonlinear shapes, capturing interactions between EEG background severity and clinical context that a straight line could never represent. The near-perfect separation suggests that the combination of EEG background features and clinical variables carries a strong, reproducible signature of acute brain injury, one that skilled human readers may perceive only partially or inconsistently.

For the multiclass task, logistic regression took the lead, achieving the highest test accuracy of 0.77 in distinguishing unilateral from bilateral cerebral injury. That a simple, interpretable linear model outperformed more flexible ensemble methods on this harder problem is a cautionary tale in applied machine learning: with a modest sample size of 201 patients, overfitting is a constant threat, and regularized linear models often generalize more reliably than their more expressive cousins. The performance gap between the binary and multiclass tasks also makes intuitive sense. Detecting that injury exists is easier than pinpointing its spatial distribution, particularly when the ground truth depends on neuroimaging that may be obtained at varying times relative to the EEG.

Performance was evaluated with precision, recall, and AUROC, the standard triad for clinical classification studies. Precision answers the question of how often a positive prediction is correct, recall captures how many true injuries the model catches, and the AUROC summarizes discriminative ability across all possible thresholds. In a PICU setting these metrics carry very different stakes. A model with high recall but modest precision could serve as a screening alarm, flagging children who warrant urgent imaging or escalation of EEG monitoring. A model with high precision could give clinicians confidence to deprioritize testing in children it deems low risk. The reported AUROC of 0.95 places the binary classifier in territory rarely reached by clinical prediction tools in pediatric neurocritical care.

The significance of this work is best understood against the backdrop of what continuous EEG can and cannot currently offer in the PICU. Prior studies from the same group and others have shown that electrographic seizures are common in critically ill children, that they are frequently non-convulsive and invisible without EEG, and that time to EEG monitoring is independently associated with outcome. Early EEG findings after pediatric cardiac arrest correlate with neurologic prognosis, and quantitative EEG with machine learning has shown promise for predicting survival after pediatric cardiac arrest. What has been missing is a prospective, generalizable tool that works across the whole heterogeneous PICU population, not just after cardiac arrest, and that predicts injury itself rather than a downstream outcome.

The study also lands in a moment when machine learning is spreading rapidly through pediatric critical care. Recent work has deployed algorithms to predict electrographic seizures in critically ill children, to forecast pediatric mortality early in the ICU stay, and to flag ward patients at risk of deterioration through tools such as the Deterioration Risk Index. Artificial intelligence is increasingly viewed as a way to compress the lag between physiologic change and clinical recognition. The present study extends that agenda to the brain itself, arguing that the EEG, long interpreted qualitatively by experts who are in short supply, can be quantified and interrogated systematically at scale.

Important caveats remain. The cohort comes from a single institution, the models rely on manually curated EEG background features rather than end-to-end learning from raw signals, and the multiclass accuracy of 0.77 leaves real room for improvement before lateralization predictions could change bedside decisions. The authors report that the study received no external funding, was approved by the Children’s National Hospital Institutional Review Board without a consent requirement, and that data are available to reviewers and editors on request. Still, the direction of travel is clear. If validated externally and eventually embedded in real-time EEG software, these classifiers could turn every routine PICU EEG into a continuous injury surveillance system, alerting clinicians to acute cerebral injury hours or days before it would otherwise be suspected, when interventions still have a chance to change the trajectory of a child’s brain and the shape of the rest of that child’s life.

Subject of Research: Machine learning applied to electroencephalography and clinical features to predict acute cerebral injury in critically ill children in the pediatric intensive care unit.

Article Title: Machine Learning Using Electroencephalography Predicts Acute Cerebral Injury in the Pediatric ICU

Article References: Sansevere, A. J., Anwar, S. M., Keenan, J. S., Conley, C. R., Staso, K., & Harrar, D. B. (2026). Machine Learning Using Electroencephalography Predicts Acute Cerebral Injury in the Pediatric ICU. Neurocritical Care. https://doi.org/10.1007/s12028-026-02596-y

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02596-y

Keywords: machine learning, electroencephalography, EEG, pediatric intensive care unit, cerebral injury, brain injury, neurocritical care, support vector machine, acute brain injury, pediatric neurology, seizures, clinical prediction