New Methods and Insights About Neurological Emergencies
We are uncovering actionable insights for some of the most severe types of acute brain injury by integrating physiology, neuroimaging, and electronic health records data streams.


Machine Learning Applied to Coma Neuroscience 

We use state-of-the-art machine learning methods leveraging multimodal healthcare data with the goal to improve prediction, diagnosis, and prognostication in acute neurological emergencies. We are developing methods that can integrate time-series data (EEG, EKG, blood pressure, intracranial physiology) with electronic health record data and neuroimaging studies to uncover new insights about the mechanisms of neurological recovery after acute brain injury that can support clinical decision-making.

Convolutional neural networks can help decode which features from individual EEG snapshots are most relevant for predictions. We string together the CNN data using long short-term neural networks to capture the longitudinal evolution of EEG data during consciousness recovery from coma after acute brain injury with the goal of providing real-time brain injury severity predictions (MRI: purple indicates more severe injury).


Coma & Critical Care Neurophysiology (Arrest, TBI, and Stroke)

We employ invasive and non-invasive brain monitoring for critically ill patients with coma, cardiac arrest, traumatic brain injury, stroke, epilepsy, or those at risk for neurological deterioration. We focus on continuous EEG monitoring modeling as well as combining scalp EEG data analysis with intracranial EEG (ECoG), cerebral blood flow, brain oxygenation, and intracranial pressure trends. Our group also has experience delivering auditory task experiments in comatose subjects in the critical care environment to identify if neural responses to these tasks predict the likelihood of consciousness recovery.

Neural State Dynamics Predict Acute Coma Recovery


t-SNE map of more than 50,000 hours of EEG data for 1,038 comatose cardiac arrest patients: different EEG signatures are associated with various degrees of brain edema severity on MRI. Each dot represents 6h of EEG data and the distance between dots indicates how different the EEG data within every 6 hours are from each other. The limited overlap between clusters indicates that there are well-defined EEG phenotypes (each color represents an EEG phenotype). Coma recovery likelihood varies across phenotypes and patient transitions between phenotypes are predictive of coma recovery.


Chord diagrams demonstrating longitudinal neural state transitions for 1,038 acutely comatose cardiac arrest patients (EEG based). Transition dynamics in the first 84h after cardiac arrest varies between patients recovering from a coma (left) and those with poor neurological outcomes (right). Patients who recover from coma are primarily stationary on higher entropy states, but the transition to and from seizure (green and orange) and burst suppression (blue) states can be observed in specific time windows. The raw EEG and spectrogram shown are representative of a high entropy state (left) and burst suppression and seizure state (right).


Spreading Depolarizations in Traumatic Brain Injury

Traumatic brain injury can trigger spreading depolarizations (three spreading depression events are shown above traveling across five intracranial electrodes). This is a massive neurophysiology phenomenon with ultra-slow brain waves that are associated with worse outcomes and are likely a source of secondary injury to the brain. These waveforms can only be recorded with invasive brain monitors and special amplifiers.


Detecting Responses to Sound and Music In Comatose Patients

Neural processing of auditory responses to novel stimuli (red) versus deviant (blue) and standard (gray) sounds in an epilepsy patient with multiple intracranial depth electrodes placed. The data shown represent electrodes targeting the left superior temporal gyrus (LAT). The brain tends to have a more pronounced response, i.e. higher amplitude, to salient complex sounds (e.g. telephone ringing) than low complexity tones (e.g. white noise). These higher-order neural responses can also be identified in non-invasive surface EEG for patients in a coma who recover consciousness.


The opioid crisis and fentanyl use have caused an increase in deaths and cardiac arrests associated with overdose. We have shown that cardiac arrest in the setting of opioid and fentanyl use is associated with more severe brain injury and that this injury might affect different brain regions compared to cardiac arrests not associated with substance use.