Neurology produced the device that created a regulatory category. The first computer-aided triage authorization in the United States was a stroke tool, and the reason is worth understanding because it explains what AI is genuinely good for in this specialty.

In large vessel occlusion stroke, the interval between imaging and intervention is the outcome variable. Not the accuracy of the read ... the time. A device that compresses that interval improves care without ever being a better reader than the radiologist who eventually reviews the study.

Stroke triage: a speed problem

The mechanism is parallel notification. Software analyzes the CT angiogram as it arrives, and if it suspects a large vessel occlusion it pushes an alert to the interventional team at the same moment the study enters the radiologist's queue. Nothing is removed from the queue. Nothing is marked on the image.

Before this existed, there was no regulatory classification for software whose entire clinical claim was speed of notification rather than accuracy of interpretation. The FDA authorized it through De Novo and wrote the special controls that every triage device since has cited as a predicate.

What the evidence actually shows

The reported time savings come largely from transfer-network studies ... spoke hospitals moving patients to thrombectomy-capable hub centers. That is where coordination delay is greatest and where parallel notification has the most to compress.

A single-site hospital with an in-house interventional team and a radiologist reading promptly will not reproduce those numbers, and should not budget as though it will.

Sensitivity is also considerably better for proximal occlusions than distal ones, and the indication is for suspected occlusion rather than for excluding it. A negative result is not a clearance to stand the team down.

Brain volumetry: a measurement problem

This is a different category entirely, and a more interesting one. Radiologists describe brain atrophy qualitatively and inconsistently ... mild, moderate, advanced for age ... and clinically meaningful year-over-year change sits well below what visual inspection can detect.

Automated volumetry is not competing with expert reading here. It is measuring something that was not being measured, which is a much stronger position than trying to outperform a radiologist at a task radiologists are good at.

The caveat is technical rather than clinical. Volumetric output is highly sensitive to acquisition parameters, and comparing scans across scanners or protocols can produce apparent change that is entirely an artifact of the equipment. Normative comparison databases may also not represent the patient's demographic, which shifts every percentile-based interpretation built on them.

EEG: a volume problem

Continuous EEG monitoring in an intensive care unit generates days of recording. No one reads it in full. That is not a criticism of anybody, it is arithmetic.

Automated seizure and spike detection is therefore not an accuracy improvement over an expert electroencephalographer. It is the only mechanism by which the data gets examined at all, which is a more honest justification than most medical AI can offer.

The cost is false positives, and in artifact-heavy ICU recordings the rate is high. Detection remains an adjunct requiring expert confirmation, and sensitivity varies substantially by seizure type ... subtle electrographic seizures being both the hardest class and frequently the most clinically important.

A related development is rapid-deployment EEG with algorithmic analysis, aimed at hospitals that cannot obtain an EEG at three in the morning. Non-convulsive status epilepticus requires EEG to diagnose and has a short treatment window, so availability rather than interpretation is the actual bottleneck.

The pattern across all three

None of these tools is trying to out-diagnose a neurologist. One compresses time, one measures the unmeasured, one makes unread data readable. That is a fair description of where clinical AI currently earns its place, in neurology and generally.

See the neurology device records and the neurology specialty hub.