Cognitive Bias in EEG Interpretation

The Bayesian skeleton of a read

Reading an ambiguous EEG is, formally, an act of Bayesian inference, and making that structure explicit is the most powerful single tool for understanding both correct interpretation and its characteristic failures. The reader begins with a prior probability that the patient has the condition in question - shaped by age, sex, referral reason, and especially the clinical history written on the requisition - and updates that prior with the likelihood that the observed waveform would occur under each competing hypothesis, arriving at a posterior probability that drives the report. This is not a loose analogy. It is exactly how a calibrated reader should behave, because the same waveform genuinely is more likely to be epileptiform in a patient with stereotyped focal seizures than in an asymptomatic volunteer, and a rational reader ought to weight it accordingly. The difficulty - the entire problem of cognitive bias - is that in practice the priors are largely implicit, often poorly calibrated, and smuggled in through the history rather than reasoned about openly. The legitimate Bayesian machinery thereby mutates silently into bias. Cognitive bias in EEG is best understood not as the presence of priors, which are correct and indispensable, but as their miscalibration and uncontrolled, invisible influence.

Likelihood ratios and a worked example

Bayes' theorem is cleanest in odds form, which is also how clinicians can actually use it at the bench. The rule is simple: posterior odds equal prior odds multiplied by the likelihood ratio. The likelihood ratio (LR) of a test result summarizes how many times more likely that result is in people with the condition than in people without it. A positive result has LR+ equal to sensitivity divided by one-minus-specificity; a negative result has LR- equal to one-minus-sensitivity divided by specificity. By the widely cited convention attributed to the Centre for Evidence-Based Medicine, the strength of a positive result climbs through rough bands: an LR+ above 10 generates a large, often decisive increase in the post-test probability, values from about 5 to 10 a moderate increase, values from about 2 to 5 only a small increase, and values below 2 a negligible one, with an LR of exactly 1 adding nothing at all. The mirror-image bands apply to negative results: an LR- below 0.1 produces a large decrease in probability and strongly helps rule a condition out, 0.1 to 0.2 a moderate decrease, and 0.2 to 0.5 only a small one. These bands are heuristics, not laws, but they give the clinician an intuition for how much any given finding should move belief, and they make clear that most single EEG findings fall in the small-to-moderate range rather than the decisive one.

Consider a deliberately hypothetical, illustrative example - the numbers are invented for teaching and are not measured values from any study. Suppose, for the sake of arithmetic, that a particular borderline sharp transient has a likelihood ratio of about 5 for the presence of epilepsy: it is five times more common in people with epilepsy than in people without. By the bands above, an LR of 5 sits right at the boundary between small and moderate evidence - deliberately chosen, because most real EEG findings live in exactly this unimpressive middle range rather than at the decisive extremes. Now apply it in two patients. In a young adult referred after a single faint with an otherwise unremarkable history, assume a low pretest probability of epilepsy of about 5 percent. Converting to odds gives roughly 0.05 to 0.95, or about 1 to 19; multiplying by the LR of 5 gives posterior odds of about 5 to 19, which converts back to a posterior probability of only about 21 percent. The same finding in a patient with several stereotyped focal-onset episodes, where we might assume a pretest probability of 60 percent, starts from odds of 60 to 40, or 1.5; multiplying by 5 gives 7.5, a posterior probability of about 88 percent. The identical waveform, carrying the identical likelihood ratio, leaves one patient probably without epilepsy and the other probably with it. Nothing about the trace changed; only the prior did. This single calculation, with its frankly hypothetical numbers, is the quantitative heart of why context-free reading is irrational and why the overdiagnosis of epilepsy is mathematically inevitable when sensitive interpretation meets low-prior patients.

The named biases and how they operate

The general distortions catalogued by cognitive psychology take specific, recognizable forms at the EEG bench, and each maps cleanly onto the Bayesian skeleton. Anchoring is excessive weight on the first piece of information encountered - typically the clinical question or a prior report - so that a requisition reading rule out epilepsy primes the reader to find epileptiform activity and to interpret every borderline transient in that light; in Bayesian terms the prior is fixed too strongly and later evidence is under-weighted. Confirmation or expectation bias is the selective search for and over-weighting of evidence that supports the leading hypothesis while discounting evidence against it; the well-replicated empirical finding that the same EEG is reported differently depending on the clinical history supplied is confirmation bias made visible, and in Bayesian terms the likelihood itself is evaluated selectively to favor the expected answer. Availability is over-estimating the probability of diagnoses that come easily to mind - the fellow who read three nonconvulsive status cases this week sees the next ambiguous record as status - which inflates the prior by ease of recall rather than true base rate. Framing is sensitivity to how information is presented rather than its content, so that possible subtle seizures, please assess and artifact-laden study steer the same data toward opposite conclusions by assigning different effective priors through wording alone. Satisfaction of search is the premature cessation of analysis once one finding is made, so that a reader who localizes a temporal sharp wave stops scanning and misses a second independent focus or a coexisting artifact - the posterior accepted before all the evidence has been incorporated.

These biases are not failures of effort or intelligence; they are the predictable behavior of a fast, pattern-completing perceptual system - the gestalt System 1 of the pattern-recognition module - operating without analytic checks. This is precisely why exhortations to try harder or be more careful are nearly useless against them, a finding that holds across the entire diagnostic-error literature. The fast system is doing exactly what it evolved to do: complete the pattern toward what is expected, quickly and automatically, below the threshold of awareness. Telling it to be more careful does not change its architecture. Effective debiasing therefore changes the process and the information available to the fast system, not the reader's willpower - a principle that determines which interventions actually work.

Base rates and the overdiagnosis of epilepsy

The most important quantitative lesson in this pillar is the behavior of base rates under an imperfect test, the phenomenon the worked example above made concrete. Even a good test produces many false positives when applied to a low-prevalence population, because the false positives are drawn from the large pool of unaffected people while the true positives are drawn from the small pool of affected ones. EEG is a vivid case. The interictal EEG has only modest sensitivity for epilepsy on a single routine study and a non-trivial false-positive rate, since benign variants and over-interpreted sharps generate epileptiform-looking findings in people without epilepsy. When such a test is applied broadly to patients with a low pretest probability - a single faint, an atypical spell, a vague screening request - a meaningful share of the positive reads will be false, exactly as the hypothetical 21 percent posterior demonstrated. Each false positive then risks the durable wrongful diagnosis catalogued in the pitfalls module. This is the mathematical engine behind the overdiagnosis of epilepsy: a fixed false-positive rate becomes clinically dominant when the prior is low, and over-reading compounds the effect by inflating that false-positive rate still further. The base-rate neglect that drives it - attending to the vividness of the waveform while ignoring the prior prevalence - is itself one of the most robust findings in the psychology of judgment.

Debiasing: changing the process, not the willpower

Because biases are properties of the process rather than of character, the interventions that work are structural. Blinded or history-masked review - forming a first impression before reading the clinical history, or having a portion of the record over-read by someone unaware of the leading diagnosis - directly attacks anchoring and confirmation by denying the fast system the prior that would distort it; the reader then incorporates the history deliberately, as an explicit second step, restoring its legitimate Bayesian value. Structured, standardized terminology, such as the ACNS critical-care EEG framework, reduces the latitude in which bias operates: by forcing a main term plus explicit modifiers rather than a single loaded label, it constrains the reader to describe what is present before interpreting it, and it makes inter-reader disagreement measurable rather than hidden. Explicit consideration of alternatives - deliberately asking what else this pattern could be, and what a benign variant would look like here - counters satisfaction of search and confirmation by forcing the likelihood to be evaluated for competing hypotheses, not just the favored one. Independent double reading and routine feedback on discordant cases recalibrate priors over time, the same deliberate-practice mechanism that builds gestalt expertise. And simply stating uncertainty in the report - recording a finding as equivocal rather than forcing a binary - prevents a tentative impression from hardening into a false anchor for the next reader, breaking the self-reinforcing cycle at its source.

These strategies have limits worth naming honestly, because overselling them is its own error. Blinding is costly and discards the genuine diagnostic value that an accurate history provides; it is best used as a structured first-pass step rather than a permanent veil, since a reader who never sees the history sacrifices the legitimate prior along with the illegitimate one. Structured terminology constrains description but cannot by itself supply the missing morphologic discipline: a reader who calls every sharp transient a discharge will simply do so in standardized language, with the bias intact beneath the formatting. Checklists and forced alternatives slow the read and can induce the very fatigue that degrades vigilance. The realistic goal is therefore not the elimination of bias - impossible for any system built on priors, and undesirable even if possible, since priors are correct - but its containment: making priors explicit, calibrating them against base rates, and building processes that force the fast system to pause precisely where the cost of its errors is highest. Debiasing is risk management, not exorcism.