Diagnostic Pitfalls

Three failure modes, one inverse problem

Because the scalp EEG is a blurred projection read under irreducible uncertainty, interpretive error is not an occasional accident to be engineered away but a structural feature of the task itself. This reframing matters: a clinician who believes error is merely a sign of insufficient care will respond to mistakes with exhortation and shame, whereas one who understands error as structural will respond with process - checklists, double reading, standardized terminology, calibration against base rates. The errors of EEG interpretation cluster into three recognizable families. Over-reading calls something epileptiform or ictal that is not, usually a benign variant or an artifact mistaken for pathology. Underrecognition misses real pathology that is present but subtle - the unobtrusive electrographic seizure, the nonconvulsive status hiding in an obtunded patient, the asymmetry overlooked on a hurried read. Misclassification assigns the wrong category to a real signal, most commonly labeling artifact as cerebral or mistaking one pattern for another. Each family has a characteristic mechanism and, decisively, a characteristic cost, and the costs are not symmetric. The central skill of this module is learning to manage the trade-off among the families rather than imagining one can simply abolish all error at once.

Over-reading and the manufacture of epilepsy

Over-reading of interictal epileptiform discharges is the best-documented and arguably most consequential of all EEG errors. The mechanism is a quiet loosening of the morphologic criteria that define a genuine discharge: a transient is called epileptiform simply because it is pointy, without the reader demanding the full constellation that the literature has tried to codify - a paroxysmal transient standing clearly out from the background, an asymmetric limb, a duration in the appropriate range, a physiologic field spanning more than a single electrode, an after-going slow wave, and a disruption of the surrounding activity. Modern criteria sets, such as those formalized by the International Federation of Clinical Neurophysiology, attempt to make these requirements explicit precisely to combat the slide toward pointy equals epileptiform. Benign variants are, by their very nature, the entities that satisfy the sharpness criterion while failing the others. Wicket waves, small sharp spikes (benign epileptiform transients of sleep), 14-and-6 positive bursts, rhythmic temporal theta of drowsiness (the pattern formerly called psychomotor variant), and SREDA (subclinical rhythmic electrographic discharge of adults) are the classic traps, each sharp or rhythmic enough to seduce a reader who has substituted pointy for epileptiform.

The empirical reality is that even experts disagree substantially about whether an individual transient is epileptiform. A large multicenter study in which experienced readers independently annotated many thousands of candidate discharges found that chance-corrected agreement on individual interictal epileptiform discharges sat only at the fair-to-moderate border - the investigators characterized it as fair - well below the level one might hope, even while agreement on whether an entire record contained any epileptiform activity was substantially better. The disagreement was largely explained by different readers applying different thresholds to a shared underlying judgment - some readers simply have a lower bar for calling a transient a spike. This finding, examined again in the evidence module, is the quantitative shadow of the over-reading problem: the boundary between epileptiform and not is genuinely blurry, and where an individual reader places it determines how often they manufacture findings.

The harm of over-reading is durable and self-reinforcing in a way that distinguishes it from most diagnostic errors. A false epileptiform read can anchor a diagnosis of epilepsy in a patient whose paroxysmal events were actually syncope, psychogenic nonepileptic events, migraine, panic, or a parasomnia. That label then exposes the patient to years of antiseizure medication with its cognitive, teratogenic, dermatologic, and metabolic burdens; to driving and occupational restrictions that can cost employment and independence; to the psychological weight of a chronic neurological diagnosis and the stigma that accompanies it; and, perversely, to confirmation of the error, because the original report becomes a written prior that biases every subsequent reader who knows the patient carries a seizure diagnosis. Studies of patients referred to tertiary epilepsy centers for medically refractory seizures have repeatedly found that a substantial minority never had epilepsy at all, and an over-interpreted EEG is consistently among the commonest reasons the misdiagnosis was made and then sustained for years. The discharge that was never there is, in this precise sense, one of the most expensive errors in all of clinical neurophysiology.

Underrecognition of subtle seizures and NCSE

The opposite failure - missing real pathology - is most dangerous in the critically ill. Nonconvulsive seizures and nonconvulsive status epilepticus (NCSE) are common in comatose, post-cardiac-arrest, and post-convulsive patients and are, by definition, invisible on examination; they declare themselves only on EEG, and often only on continuous EEG, because a routine twenty-minute study can fall entirely between events. Multiple critical-care cohorts have shown that a meaningful fraction of nonconvulsive seizures are detected only after many hours of monitoring and would have been missed by a short record. Underrecognition in this setting has several distinct mechanisms. The pattern may be intrinsically subtle - low in amplitude, slow to evolve, or buried in an ICU record saturated with ventilator, chest-physiotherapy, dialysis, and infusion-pump artifact. It may sit squarely on the ictal-interictal continuum, where frequent or rhythmic periodic discharges are neither clearly ictal nor clearly interictal and require correlation with the examination, sometimes a trial of a benzodiazepine, and continued monitoring to resolve. Or the reader may simply not look long enough, sampling minutes when the diagnosis demanded hours.

Underrecognition also afflicts routine outpatient EEG, though usually less catastrophically. A genuine focal discharge can be dismissed as artifact, a subtle asymmetry of the posterior dominant rhythm can be overlooked, an electrodecremental onset can be read as movement, and a record obtained only in wakefulness can miss discharges that would have appeared in sleep or after sleep deprivation. The unifying lesson, which deserves to be stated as a principle, is that absence of a finding on a short or hastily reviewed record is weak evidence of absence. The diagnostic yield of EEG rises with sleep, with prior sleep deprivation, with repeated studies, and with longer monitoring - which is precisely why these activation and duration strategies exist and are recommended. A single normal routine EEG never excludes epilepsy; sensitivity for interictal discharges on one routine study is modest, and serial or prolonged recording raises it substantially. A negative short ICU record never excludes nonconvulsive status. The clinician who reports a normal study has not excluded disease but has merely failed to find it within the window examined, and the report should say so.

Misclassification: when artifact wears the costume of brain

Misclassification of artifact as cerebral activity is, across all settings and all levels of experience, the commonest single cause of EEG error. The defense against it is field logic, the same constraint that anchored the pattern-recognition module. Genuine cerebral activity volume-conducts to neighboring electrodes in an anatomically sensible, spatially smooth pattern dictated by physics; artifact typically violates that logic - confined to a single electrode, or distributed in a way that ignores anatomy, or time-locked to a demonstrably non-cerebral source. Each common artifact carries a recognizable signature. ECG artifact betrays itself by its lock-step relationship to the QRS complex on a simultaneously displayed ECG channel. Pulse artifact reveals a fixed delay after each heartbeat at an electrode lying over a scalp artery. Muscle artifact declares itself by its very high frequency content and its disappearance when the patient relaxes. Eye-movement artifact shows frontal predominance and obeys the direction of gaze, with characteristic out-of-phase deflections at the two frontopolar electrodes on lateral movement. Electrode-pop artifact appears as an abrupt single-channel step with no field at all. The corollary, and the practical heart of this module, is that the most powerful diagnostic instrument in EEG is not a filter but a reflexive question - does this have a field that makes anatomic sense? - asked and answered before any label is assigned. Filters can suppress an artifact's appearance, but only field logic can establish its nature, and a reader who reaches for the filter before the question has skipped the only step that distinguishes signal from noise.

The asymmetry of error costs

Every interpretive threshold trades one kind of error against another, and the right threshold depends on which error costs more in the specific clinical context. This is the conceptual bridge to the Bayesian decision theory of the next two modules, but it can be stated plainly here. The two dominant errors - the false positive of over-reading and the false negative of underrecognition - are not equivalent, and their relative weight shifts dramatically with setting. In the ambulatory evaluation of a single transient event in an otherwise well person, over-reading is usually the more harmful error, because a wrongful epilepsy diagnosis is durable, exposes the patient to real and prolonged harm, and propagates through the record as a contaminating prior. Here the prudent reader holds a deliberately high bar and is entirely comfortable reporting an EEG as normal or nonspecific, accepting that they will occasionally miss a genuine discharge in exchange for rarely manufacturing a false one. In the comatose ICU patient with possible nonconvulsive status, the balance often inverts: missing ongoing seizures can permit continued neuronal injury and worsen outcome, so a lower threshold for suspicion, a longer recording, and an explicit treatment trial may be justified, accepting that some periodic patterns will be treated that might not have required it.

The same waveform can therefore warrant different actions depending on the pretest probability and on the asymmetry of consequences - a stance that is not inconsistency but the correct application of decision theory. A formal way to express this, developed quantitatively in the cognitive-bias module, is that the optimal decision threshold should shift with the ratio of the cost of a false positive to the cost of a false negative, weighted by the prior odds of disease. When false positives are catastrophic and disease is rare, the bar should be high; when false negatives are catastrophic and disease is plausible, the bar should be low. The clinician who applies a single fixed threshold to every record, regardless of setting, is guaranteed to be miscalibrated in at least one of these contexts. The honest report acknowledges this by stating uncertainty explicitly rather than forcing every finding into a binary normal-or-abnormal verdict that the underlying evidence cannot support.