ICU EEG Applications

In the intensive care unit the EEG changes its job. It is no longer chiefly a tool for diagnosing epilepsy but a continuous monitor of brain function in patients who cannot be examined - the comatose, the sedated, the pharmacologically paralyzed. Continuous EEG (cEEG) in this setting answers four questions that the bedside neurologic exam cannot: how severely is the brain depressed, is the brain seizing invisibly, what is the trajectory of recovery, and how much of what we see is the drug rather than the disease. These questions are tightly coupled, and the great error of ICU electroencephalography is to answer one while forgetting the others - most commonly, to read severity or prognosis off a tracing without accounting for sedation. Every section of this module returns to that coupling, because the identical waveform can mean opposite things depending on the pharmacologic and physiologic context in which it was recorded.

Grading the comatose brain

The depth and quality of coma are read from a small number of robust features. Continuity is foremost: a continuous background is more reassuring than a discontinuous one, which is in turn better than a burst-suppression pattern (bursts of activity alternating with intervals of flat suppression), which is better than complete suppression. Reactivity - any reproducible change in the EEG in response to an external stimulus such as noise, light, or noxious touch - is arguably the single most important bedside-testable feature, because its presence implies functioning cortical-subcortical networks and carries favorable prognostic weight, while its absence is ominous. The reader also grades variability (does the record evolve over minutes and hours, or is it monotonous?) and the dominant frequency (faster is generally better than a record dominated by slow delta). A continuous, reactive, variable background is the favorable end of every one of these axes; a suppressed, unreactive, invariant background is the grave end.

Specific patterns layer onto these axes, and the field's central achievement is a shared vocabulary for them: the ACNS Standardized Critical Care EEG Terminology, whose current edition dates to 2021 and is the third in a lineage running from the original 2005 proposal through the 2012 revision. Its power is that it decomposes any periodic or rhythmic pattern into independent, agreed components. Main Term 1 specifies location - generalized, lateralized, bilateral independent (an independent pattern in each hemisphere), unilateral independent (two independent patterns in the same hemisphere), or multifocal. Main Term 2 specifies the pattern itself - periodic discharges (PDs), rhythmic delta activity (RDA), or spike-and-wave or sharp-and-wave (SW). A set of modifiers then captures prevalence, duration, frequency, sharpness, amplitude, polarity, evolution, and fluctuation, along with whether the pattern is stimulus-induced or stimulus-terminated. The 'plus' modifiers note superimposed features and are pattern-specific: +F for superimposed fast activity (applicable to both PDs and RDA), +R for superimposed rhythmic activity (PDs only), and +S for a superimposed sharp or spike component (RDA only). This terminology exists so that different readers describe the same pattern the same way and so that ambiguous, potentially ictal patterns can be tracked over time; without it, the most important judgment in ICU EEG - is this pattern a seizure? - cannot even be communicated.

It is worth dwelling on why this descriptive discipline, rather than a single diagnostic label, is the right tool for the critically ill brain. The patterns that dominate the ICU - periodic discharges, rhythmic delta, fluctuating runs of sharp activity - sit on a spectrum of epileptogenicity rather than falling into clean categories of seizure and non-seizure, and the same morphology carries different risk depending on its frequency, its evolution, and the clinical context. A standardized description allows the relevant risk to be attached to the specific pattern: lateralized periodic discharges, for example, are associated with a substantially elevated probability of accompanying or subsequent electrographic seizures, whereas generalized periodic discharges of metabolic origin may not be. The terminology also makes change over time legible, which is the heart of monitoring - a pattern that is accelerating in frequency, acquiring a plus modifier, or beginning to evolve spatially is moving toward the ictal end of the spectrum, and that trajectory is often more informative than any single epoch. The reader's task in the ICU is therefore less to name a disease than to characterize a dynamic process and to communicate its evolving danger to the team managing the patient hour by hour.

Detecting nonconvulsive seizures

The most actionable contribution of cEEG is the detection of nonconvulsive seizures (NCS) and nonconvulsive status epilepticus (NCSE) - electrographic seizures with little or no overt motor manifestation, invisible to the bedside exam by definition. They are common in the critically ill, particularly after convulsive status epilepticus, in acute brain injury (traumatic, hemorrhagic, ischemic), and in unexplained coma. The landmark cohort by Claassen and colleagues, of 570 critically ill patients, found electrographic seizures in 19 percent, of which 92 percent were exclusively nonconvulsive - and its data on timing are frequently misquoted, so they are worth stating precisely. Across all patients who would eventually seize, the first seizure appeared within the first 24 hours in 88 percent. The clinically decisive split is by level of consciousness: the first seizure was recorded only after more than 24 hours in 20 percent of comatose patients versus 5 percent of non-comatose patients (odds ratio about 4.5), and coma was itself the strongest predictor of having any seizure (odds ratio about 7.7).

The practical rule that follows is therefore risk-stratified, not one-size-fits-all. A non-comatose patient whose initial recording shows no epileptiform abnormalities has a low and rapidly falling residual seizure risk - by roughly 24 hours the cumulative risk has dropped below about 5 percent in the later work of Struck and colleagues - whereas a comatose patient, especially one with prior seizures, carries a high enough incidence that monitoring frequently must extend beyond 24 hours and often to 48 or more to reach acceptable detection. The contemporary standard is thus to record at least 24 hours in patients with unexplained altered consciousness or persistent encephalopathy after a clinical seizure, and longer in the comatose, where the cumulative detection of seizures continues to climb with time. This supersedes any blanket prescription to monitor every patient for a fixed 48 hours: the duration is set by the patient's risk and by what the early record shows.

The very fact that detection yield climbs with recording duration is the argument for continuous rather than spot EEG in this population, and it reframes the clinical question from a single read into an exercise in sequential probability. Each additional hour of seizure-free recording in a low-risk patient progressively lowers the residual probability that a seizure will appear, while in a high-risk patient - the comatose, those with prior seizures, those with periodic discharges on the initial epochs - the prior is high enough that a negative early record is not yet reassuring. The operational consequence is that monitoring duration should be titrated to risk and to what the record shows, and that the burden of reviewing days of multichannel EEG is managed with quantitative trending tools that compress the data without replacing the raw signal. This sequential, risk-stratified logic is the modern descendant of the older blanket recommendations, and it depends on the same Bayesian discipline that governs the rest of electroencephalography: the meaning of a finding, or of its absence, is conditional on the pre-test probability the clinical situation supplies.

The interpretive difficulty is the ictal-interictal continuum (IIC): periodic and rhythmic patterns that are neither clearly seizures nor clearly benign. The 2021 ACNS terminology gives the IIC a formal definition - a pattern qualifies if it does not meet electrographic-seizure criteria but is a PD or SW averaging more than 1.0 and up to 2.5 Hz, or a PD or SW averaging 0.5 to 1.0 Hz with a plus modifier or fluctuation, or lateralized RDA averaging more than 1.0 Hz with a plus modifier or fluctuation. Two precision points follow: the low-frequency arm is bounded at 0.5 Hz, not open-ended, and the RDA arm requires lateralized RDA, because generalized RDA alone does not qualify. Critically, there is no numeric score that places a pattern at a point along the continuum; the IIC is a binary electrographic qualifier, and what is actually quantified is the seizure risk that different patterns carry - lateralized periodic discharges, for instance, carry a high associated seizure risk. The IIC is a description that warrants a diagnostic trial and continued monitoring, not a diagnosis in itself.

When a periodic or rhythmic pattern raises the question of NCSE, the Salzburg criteria provide the operational framework. In a patient without a known epileptic encephalopathy, epileptiform discharges exceeding 2.5 Hz (persisting at least 10 seconds) meet criteria on their own. Discharges at or below 2.5 Hz, or rhythmic delta, require one additional criterion: subtle clinical ictal phenomena, typical spatiotemporal evolution, or significant improvement after intravenous antiseizure medication. The grading of certainty is essential and often dropped: the benzodiazepine trial counts toward definite NCSE only when both clinical state and EEG improve; EEG improvement without clinical improvement yields only possible NCSE. Quantitative EEG trends - color density spectral arrays and amplitude-integrated displays - compress hours of data into a single screen and are invaluable for screening and for spotting the evolution that marks a seizure, but every quantitative flag must be confirmed against the raw tracing, because artifact readily mimics the IIC on a spectrogram.

Post-arrest prognostication and the self-fulfilling prophecy

After cardiac arrest, EEG is one pillar of multimodal prognostication, never used in isolation - and the reason it must never stand alone is the danger of a self-fulfilling prophecy, in which a pessimistic prediction leads to withdrawal of life-sustaining therapy that then guarantees the poor outcome the prediction foresaw. The 2021 ERC/ESICM guidelines are built around mitigating exactly this bias. Certain patterns carry strong but not absolute weight: in the Westhall terminology, the highly malignant EEG patterns are suppression (background under about 10 microvolts) and burst-suppression, recorded after rewarming and off confounding sedation, typically at high specificity but only moderate sensitivity. Reactivity, once treated as a primary criterion, was demoted to a modifier because its interrater agreement is only fair, but an unreactive highly malignant background remains more specific for poor outcome than a highly malignant background alone. Conversely, the recovery of a continuous, reactive background within the first day or two is among the more favorable EEG signs.

The cardinal discipline is timing and confounders. Highly malignant patterns are assessed more than 24 hours after arrest, after rewarming and after sedative clearance, and the full multimodal algorithm is applied at 72 hours or later in patients who remain comatose with a motor score of 3 or less. To guard against the self-fulfilling prophecy, the guidelines require at least two concordant predictors from among bilaterally absent pupillary and corneal reflexes, bilaterally absent N20 somatosensory evoked potentials, an unreactive highly malignant EEG, markedly elevated neuron-specific enolase, status myoclonus (the sustained variety, not isolated early myoclonus), and diffuse anoxic injury on CT or MRI. No single test, EEG included, may be used alone to predict a poor outcome. A further nuance comes from the TELSTAR randomized trial, which tested aggressive antiseizure treatment of rhythmic and periodic patterns after arrest and found no benefit - poor outcomes were essentially identical with and without aggressive suppression. The lesson is specific: in the post-anoxic population, suppressing rhythmic and periodic patterns for their own sake does not improve outcome, which is distinct from the separate obligation to treat clear electrographic seizures arising from other causes.

Sedation is the great confounder of all ICU EEG, not only of prognostication, and its effects are dose-dependent and recapitulate the entire spectrum of pathological depression. The GABAergic agents - propofol, the benzodiazepines, and the barbiturates - produce, with increasing dose, fast beta activity, then slowing and disorganization of the background, then discontinuity, and ultimately burst-suppression and isoelectric suppression. (The low-dose beta is in part the paradoxical-excitation phenomenon of disinhibition.) Propofol and high-dose midazolam are titrated to burst-suppression deliberately in the treatment of refractory and super-refractory status epilepticus, so the same pattern that signals catastrophe in an unsedated post-arrest patient is a therapeutic target in another - although it is worth flagging that burst-suppression is the traditional target rather than an evidence-mandated one, since it has not been shown superior to simple seizure cessation for outcome.

Two agent-specific exceptions are essential to avoid misreading. Ketamine, an NMDA antagonist, does not follow the GABAergic progression and does not produce burst-suppression; it generates a characteristic alternation of slow delta with fast beta-gamma activity, roughly in the 25 to 40 Hz range. And the frontal shift of coherent alpha known as anteriorization - a hallmark sometimes loosely attributed to general anesthesia in the aggregate - is specifically a feature of the GABAergic anesthetics such as propofol and the volatile ethers; ketamine and the alpha-2 agonist dexmedetomidine do not produce it, with dexmedetomidine instead generating a spindle-rich, NREM-like pattern that reflects arousable sedation rather than propofol-type unconsciousness. The expert reader therefore never interprets an ICU EEG without first asking what drugs are running and at what dose, because the identical tracing means opposite things depending on the answer.

The recurring theme across coma grading, seizure detection, and prognostication is that the EEG is a powerful but incomplete instrument whose value is realized only in combination. In neuroprognostication this is formalized as the requirement for concordant predictors, but the same logic applies throughout the unit. An EEG that suggests a poor prognosis acquires its real weight only alongside the neurological examination - the pupillary and corneal reflexes, the motor response - and alongside the somatosensory evoked potentials, whose bilaterally absent N20 cortical response is among the most robust single markers of severe cortical injury, and alongside structural imaging and biochemical markers such as neuron-specific enolase. The discipline of not over-reading a single modality is, in the end, the same discipline that governs the whole field: an electrographic pattern is evidence that shifts a probability, and the responsible clinician integrates it with the rest of the evidence rather than allowing one tracing - especially one recorded under sedation or hypothermia, or too early after an arrest - to drive an irreversible decision. The cost of forgetting this in the ICU is uniquely high, because the decision in question is frequently the withdrawal of life-sustaining therapy.