Ceribell

The hardware exists to feed the algorithm. A conventional EEG requires a technologist to place electrodes and a neurologist to read the tracing, which is exactly the bottleneck the company removes: Clarity is a machine learning algorithm interpreting EEG every ten seconds across all channels to produce a seizure burden estimate that a bedside clinician can act on without a neurologist present. The headband is a delivery mechanism for that interpretation, and the FDA clearances attach to the algorithm.

The system alerts rather than treats, and it is designed for a specific oversight gap: emergency departments and ICUs where 24/7 neurology EEG interpretation is unavailable. It provides continuous monitoring with instantaneous bedside alerts for suspected status epilepticus, plus a remote portal so a neurologist can access the real time tracing and guide treatment.

That combination, an actionable bedside signal with an expert able to look at the underlying data remotely, is a well constructed oversight design for an acute setting. Buyers should still establish local protocol for who confirms an alert before anticonvulsant escalation.

The algorithm's operation is described with unusual specificity: a cloud based machine learning model interpreting all EEG channels every ten seconds, producing seizure burden defined as the amount of seizure activity in the prior five minute window, with pre annotations to support efficient human EEG interpretation. Defining the output metric explicitly rather than emitting an opaque risk score is good practice. Validation dataset sizes are disclosed for the regulatory submissions. Model architecture and per class performance figures are not published.

Nothing identifies any party in the chain: no model or model family, no hosting provider, no sub processor list was located in two passes, and no stewardship or data governance disclosure was found, for a cloud based system that carries continuous neurological data across the hospital boundary by design and exposes it through a remote clinician portal.

The training question is unusually concrete here rather than speculative, and it can be established from the regulatory record rather than inferred. The company describes an extensive electroencephalography database, and its submissions were supported by data from many hundreds of patients for the neonatal indication and well over a thousand for the paediatric one, so recordings from real patients demonstrably feed algorithm development.

What is not published is whether recordings from routine clinical use at customer hospitals contribute, on what basis, what de identification is applied to a continuous physiological signal, or whether a hospital can decline. The neonatal extension changes the weight of that question rather than its shape.

Data about a neonate is durable in a way adult data is not: it concerns a person whose entire life lies ahead, who cannot consent, and whose parents are consenting during an acute crisis. Ask what is transmitted from the bedside device and retained, for how long, whether customer recordings inform development, and what applies specifically to neonatal data.

Evidence is anchored in regulatory validation at notable scale rather than published outcome trials. The pediatric clearance was supported by EEG data from more than 1,700 patients, which the company reports FDA data indicates was the largest validation dataset ever used for clearance of a seizure detection system, and the neonatal clearance was supported by data from more than 700 patients, described as the largest known neonatal validation set.

Independent literature confirms deployment across academic and community hospitals in multiple states. What would move this to an A is published evidence that faster detection changes patient outcomes rather than detection performance alone.

Converted from Not Rated. No stewardship framework or data governance disclosure was located, and the prior note's characterisation holds: a cloud based system processing continuous neurological data from critically ill patients, with a remote clinician portal, so the data crosses the hospital boundary by design.

The patient population has since widened in a way that raises the stakes. Clearances obtained during 2025 extended seizure detection to paediatric patients and then to pre term neonates, so the system now records continuous brain activity from newborns in intensive care. Data about a neonate is durable in a way adult data is not: it concerns a person whose entire life lies ahead, who cannot consent, and whose parents are consenting during an acute crisis.

The training question is unusually concrete here rather than speculative. The company describes an extensive electroencephalography database, and its regulatory submissions were supported by data from many hundreds of patients for the neonatal indication and well over a thousand for the paediatric one. Recordings from real patients therefore demonstrably feed algorithm development. What is not published is whether recordings from routine clinical use at customer hospitals contribute, under what basis, what de identification is applied to a continuous physiological signal, and whether a hospital can decline.

Ask what is transmitted from the bedside device and retained, for how long, whether customer recordings inform algorithm development and under what agreement, and what applies specifically to neonatal data.

Converted from Not Rated. No business associate statement, availability, scope or subprocessor list was located. The prior note's reasoning holds: business associate arrangements are structurally required for a cloud platform deployed in United States emergency departments and intensive care units, and the absence is one of publication rather than of practice.

Two features of this deployment make the terms worth reading rather than assuming.

The first is the remote clinician portal. The system's value is partly that a neurologist who is not present can review the recording, which means identifiable physiological data is being made available outside the treating institution's walls to a clinician who may not be its employee. Whether that reviewer is contracted by the hospital, by the vendor, or by a third party service determines who is responsible for their access, and the arrangement should be named rather than inferred from the product description.

The second is the patient population. The system is now cleared down to pre term neonates, and paediatric and neonatal records attract additional protections in several states, with longer retention obligations and different rules on parental access.

A third point applies to any device vendor. Clearance and the health privacy rule govern different questions, and holding several clearances says nothing about the adequacy of a business associate agreement. A hospital's procurement should treat the two as separate reviews.

Ask which entity signs, who the remote reviewers are and under whose authority they access recordings, what retention applies, and what happens to the accumulated recordings at contract end.

Converted from Not Rated. No commercial attestation and no trust centre were located, and the prior note was right that clearance speaks to safety and effectiveness rather than to information security. That reasoning needs one refinement, and the refinement changes the grade.

Since 2023, a submission for a networked software device must include a cybersecurity plan, a software bill of materials, and processes for monitoring and delivering post market security updates. This company has obtained several clearances since that requirement took effect, including an expanded seizure detection indication for paediatric patients in 2025, a further clearance covering pre term neonates later that year, and a delirium monitoring indication in the same period. Each of those submissions will have carried cybersecurity documentation reviewed by the regulator as a condition of reaching the market.

That is a real form of assurance and it is different in kind from an attestation. It examines this product's security design and update lifecycle rather than the organisation's control environment, it is reviewed by a body with enforcement powers, and it is not published. Vintage matters here in the vendor's favour: an older clearance would carry no such review.

A second route exists and was not exhausted in this pass. The company is listed on a United States exchange, so its annual report must describe its processes for assessing and managing cybersecurity risk, board oversight, and accountable management. That item was not retrieved, so nothing here rests on it, and it should be read before anyone relies on this row.

Ask for the cybersecurity documentation and software bill of materials, and whether any organisational attestation covers the cloud platform and clinician portal.

A clear regulatory leader in its niche, with a progression that is worth reading as a pattern. The Ceribell System is FDA 510(k) cleared for detecting suspected seizure activity, and the Clarity algorithm received successive clearances extending its cleared population: adults, then patients aged 1 and older in April 2025, then preterm neonates and older in November 2025, making it the first and only AI point of care EEG cleared across the full age range from preterm newborns through adults.

The company also reports the first FDA cleared instantaneous bedside alert for suspected status epilepticus. Expanding cleared indications incrementally with progressively larger validation datasets is the disciplined version of this pathway.

No formal governance framework was located, but the sequence of age specific clearances functions as de facto population validation: rather than claiming pediatric and neonatal performance from adult data, the company generated separate validation datasets and sought separate clearances for each population. That is the substance bias disclosure is meant to produce, applied to the demographic axis that matters most for EEG, where neonatal brain activity differs fundamentally from adult. No published analysis addresses other sources of variation.

The regulatory floor applies here and the vendor adds something to it, which is why this sits at the upper end of the band. As a cleared device manufacturer the company carries adverse event reporting, complaint handling and correction obligations that exist whether or not it advertises them, and a hospital or a clinician has a route that does not depend on the vendor granting one. What it adds is a defined output.

The algorithm's operation is described specifically, interpreting all channels every ten seconds and producing seizure burden defined as the amount of seizure activity in the prior five minute window, with pre annotations supporting human interpretation rather than replacing it.

Defining the output metric explicitly rather than emitting an opaque risk score is good practice and it is what makes the number contestable: a clinician who disagrees knows exactly what was measured over what window. Validation dataset sizes are disclosed for the regulatory submissions. Held below the top grade because no per class performance figures are published, no architecture is described, and no warranty, indemnity or remediation commitment attaches.

The population widened materially during 2025 to paediatric patients and then to pre term neonates, and no separate performance breakdown for those populations was located, which is the figure that matters most since a model developed on adults does not transfer to a neonatal brain. Ask for per population sensitivity and specificity, and for what the vendor commits to on a missed seizure.

Converted from Not Rated. No electronic health record integration, interoperability standard or named connection was located, and the prior note identified the consequence precisely: the workflow surface is the bedside device plus a remote clinician portal rather than the chart.

One detail in that note deserves to be the headline, because it is a specific and checkable clinical safety issue rather than a general interoperability complaint. Medication annotation happens inside the vendor's own interface rather than flowing from the medication record. A clinician treating a seizure administers an anti seizure medication, which is recorded in the hospital's administration record, and separately notes it in the electroencephalography interface so the recording can be interpreted against it. That is double entry at the bedside during a neurological emergency, which is the moment least suited to it, and the two records can diverge. A recording annotated with a drug that was ordered but not given, or missing one that was, will be read wrongly by whoever interprets it later, including a remote reviewer with no other view of the case.

The absence of a result path compounds it. If the interpretation does not return to the chart as a discrete finding, the next clinician sees the patient without seeing what the monitoring showed.

This is not a criticism of the product's design, which prioritises speed of deployment over integration and is right to for an emergency device. It is a question a hospital should resolve in implementation rather than discover in use.

Ask whether medication data can flow in from the administration record, and whether interpretations return to the chart in any structured form.

Explicitly cloud based, with the Clarity algorithm described as a cloud hosted model and a portal providing remote real time access for neurologists. Beyond that characterization no tenancy, hosting, or data residency terms are published, which is a live question for a system operating in critical care.

No published pricing. As a public company the vendor discloses financial results in aggregate, which gives a buyer more visibility into business health than most vendors in this index provide, but nothing about per device, per use, or subscription cost is published.

Deliberately narrow clinically and unusually broad demographically. The clinical scope is electrographic seizure detection in acute care, specifically emergency departments and intensive care units including neonatal ICUs, where the company's own framing is that up to 90 percent of NICU seizures go undetected without monitoring. Against that single indication, the cleared population now spans preterm neonates through adults, which is wider than any comparable seizure detection system. Outpatient neurology and epilepsy monitoring units are not the target setting.