Blood Flow and Brain Waves Combine to Detect Consciousness in Severely Injured Patients

Determining whether a patient with severe brain injury is conscious is one of the most consequential and difficult tasks in medicine. In the neurocritical care unit, sedation, metabolic disturbances, and the sheer severity of injury can mask awareness behind an unresponsive exterior, leaving clinicians to make high-stakes judgments from bedside examinations that are notoriously error-prone. Now, a team of researchers spanning biomedical engineering and neurology has shown that a machine learning model fed with two complementary streams of brain data—electrical activity recorded by electroencephalography and blood flow measured directly in the injured brain—can classify consciousness with substantially greater accuracy than EEG alone.

The study, published in the journal Neurocritical Care, was led by Farzad Azizi Zade, Irfaan Dar, Brandon Foreman, and Ulas Sunar, with affiliations at Ferdowsi University of Mashhad, Stony Brook University, and the University of Cincinnati. The team retrospectively analyzed data from 26 adults who had undergone multimodal neuromonitoring during their critical illness. Rather than relying on a single snapshot of brain function, the researchers segmented the continuous monitoring signals into 30-minute windows taken after each probe recalibration, ensuring that the features fed into their models reflected stable, well-calibrated measurements of the neurovascular system.

The central insight of the work is that consciousness is not purely an electrical phenomenon. EEG, the workhorse of neurocritical care monitoring, captures the synchronized firing of cortical neurons but offers only a limited view of the neurovascular unit—the tightly coupled system of neurons, glial cells, and blood vessels that sustains brain function. Cerebral blood flow, by contrast, reflects the metabolic and hemodynamic machinery that supports neural activity. The researchers hypothesized that combining these two modalities would give a machine learning classifier a richer, more robust signature of conscious brain states than either signal could provide alone.

To test this hypothesis, the team extracted a carefully curated set of features from both signals. On the EEG side, they computed band powers across the classic frequency ranges from delta through beta, along with two composite indices that have proven sensitive to states of awareness: the alpha–delta ratio, known as ADR, and the alpha divided by the sum of delta and theta power, known as ADTR, as well as total EEG power. On the blood flow side, they focused on low-frequency oscillations in perfusion, computing power in specific frequency bands including band IV spanning 0.027 to 0.073 hertz, band V spanning 0.01 to 0.027 hertz, and a broad band covering 0 to 0.5 hertz. These slow fluctuations in cerebral perfusion are thought to reflect the interplay of vascular regulation, autonomic influences, and intrinsic brain dynamics, and prior research has linked their disruption to traumatic brain injury.

The classification engine at the heart of the study was a random forest model, an ensemble machine learning method that builds many decision trees on random subsets of the data and aggregates their votes. Random forests are well suited to clinical datasets of modest size because they resist overfitting and provide a measure of feature importance. The researchers trained the model using K-fold cross-validation, a technique that repeatedly partitions the data so that every patient contributes to both training and testing, and they excluded highly correlated features—those with a correlation coefficient above 0.8—from simultaneous use, preventing redundant information from inflating performance estimates. Performance was summarized with the area under the receiver operating characteristic curve, or ROC–AUC, alongside accuracy, with confusion matrices reported for the best-performing feature combinations.

The results were striking. Multimodal feature combinations significantly outperformed models built on EEG features alone. The single best combination—EEG alpha–delta ratio, total EEG power, and two cerebral blood flow features from bands B and V—achieved a ROC–AUC of 0.86 and an accuracy of 82 percent. In clinical terms, that means the model could distinguish conscious from unconscious states in roughly four out of five monitoring windows, using information that is already being collected, or could readily be collected, at the bedside of critically ill patients. The finding that slow blood flow oscillations carried discriminative information that EEG lacked underscores the value of looking beyond electrical activity when probing the injured brain.

Perhaps the most forward-looking result came from two exploratory cases in which patients were monitored with diffuse correlation spectroscopy, or DCS, a noninvasive optical technique. DCS uses near-infrared laser light to measure blood flow in the microvasculature of the cortex by detecting fluctuations in the scattering of light caused by moving red blood cells. When the researchers applied the same feature framework to these noninvasive DCS measurements, the resulting predictions were concordant with those obtained using invasive cerebral blood flow probes. This is a crucial proof of concept: it suggests that the multimodal approach need not depend on surgically implanted monitors, opening a path toward noninvasive, real-time consciousness assessment in a far broader population of patients.

The clinical context makes the advance particularly timely. Studies of disorders of consciousness have repeatedly shown that standardized neurobehavioral assessments detect awareness in a substantial fraction of patients who appear unresponsive on routine bedside examination, a phenomenon sometimes called covert consciousness. Research in the intensive care unit has even demonstrated that some acutely injured but unresponsive patients show brain activation on task-based imaging, and that such covert cognition carries prognostic weight. Meanwhile, multimodal approaches combining FDG-PET imaging with EEG have improved both diagnosis and prognostication, and machine learning has been increasingly applied across the neuro-ICU to predict intracranial pressure crises, delayed cerebral ischemia, and surgical needs. The new study extends this trajectory by fusing hemodynamic and electrophysiological monitoring into a single classification framework tailored to the neurocritical care setting.

The technical foundations of the approach draw on a rich literature. Low-frequency oscillations in cerebral hemodynamics have been studied extensively in resting-state functional MRI and in optical imaging, and coupled oscillator models of cardiovascular and brain interactions provide a theoretical basis for why these slow rhythms carry information about brain state. Cross-frequency coupling between cerebral blood flow velocity and EEG has been documented in stroke patients, and quantitative EEG measures such as the alpha–delta ratio have long been used to track encephalopathy and recovery. By feeding these established physiological markers into a modern ensemble classifier, the researchers bridged classical neurophysiology and contemporary data science in a way that is designed to work with the noisy, artifact-laden signals typical of the intensive care environment.

The authors are careful to frame the work as a step toward, not a replacement for, clinical judgment. The study was retrospective, involved 26 adults, and used invasive monitoring data in most cases, so prospective validation in larger and more diverse cohorts will be essential before such models can inform decisions about withdrawing sedation, pursuing rehabilitation, or assessing prognosis. Nevertheless, the combination of 0.86 ROC–AUC performance, concordance with noninvasive optical monitoring, and support from an NIH Brain Initiative grant points toward a future in which bedside tools continuously integrate EEG and blood flow signals to give clinicians a real-time, quantitative read on consciousness. For the families of patients lying silently in the neuro-ICU, and for the physicians charged with deciding what those silences mean, such a tool could transform one of medicine’s most uncertain judgments into a measured, data-driven assessment of the injured brain’s capacity for awareness.

Subject of Research: Machine learning classification of consciousness in neurocritical care using combined EEG and cerebral blood flow features

Article Title: Binary Classification of Consciousness Using Cerebral Blood Flow and EEG Features

Article References: Zade, F. A., Dar, I., Foreman, B., & Sunar, U. (2026). Binary Classification of Consciousness Using Cerebral Blood Flow and EEG Features. Neurocritical Care. https://doi.org/10.1007/s12028-026-02603-2

Image Credits: AI Generated

DOI: 10.1007/s12028-026-02603-2

Keywords: consciousness, EEG, cerebral blood flow, machine learning, random forest, neurocritical care, disorders of consciousness, diffuse correlation spectroscopy, low-frequency oscillations, neuromonitoring, alpha-delta ratio, brain injury