Ceribell

The model is the engine of the bedside function, and the platform carries value without it. Clarity, a cloud machine learning algorithm, reads the EEG every ten seconds across all channels to estimate seizure burden and alert to suspected status epilepticus, and a second model scores EEG features associated with delirium; without them, a bedside clinician in an emergency department or ICU has no interpretation until a neurologist reads the tracing.

The rest of the system stands on its own: a headband placed in minutes, a recorder, and a portal where neurologists review the full EEG remotely, which is the rapid EEG acquisition the company sold before its seizure detection software was cleared.

What the algorithm does and does not do is published, with its alert threshold and the route to a person. Clarity reports seizure burden, the share of the prior five minutes with seizure activity, and alerts at the bedside to suspected status epilepticus when burden reaches 90 percent or more; the delirium score updates every ten seconds. The company states the limits directly: EEG is one component of diagnostic insight and the decision to treat should combine clinical history, presentation and objective data; the algorithm measures overall seizure burden rather than labeling individual epileptiform events; it is tuned not to miss status epilepticus and may over-call highly abnormal activity; and some disagreement with expert readers should be expected. Neurologists can review the full EEG and pre-annotations remotely in real time through the portal, with custom notifications.

A hospital should still set local protocol for who confirms an alert before escalating anti-seizure medication.

The approach and outputs are described precisely, without version or update discipline. Clarity is a cloud machine learning algorithm trained on thousands of hours of EEG annotated by epileptologists, reading all channels every ten seconds and reporting seizure burden over the prior five minutes, with pre-annotations for reviewers and an Insights Bar that flags further areas of concern; the delirium model scores quantitative EEG features associated with delirium every ten seconds. Defining the output as a measured burden over a stated window lets a clinician contest it.

The company says the algorithm is continuously improving with training on more cases, and nothing published says how versions are identified, validated or communicated to hospitals, or how that fits the cleared versions.

A structurally short chain: the algorithms are Ceribell's own and cleared as its devices. Clarity and the delirium model were built in house on expert-annotated EEG and run on the company's cloud service, so no outside model provider sits between a recording and an alert.

The parties around the data are not named: the cloud host is unstated, the neurologists reviewing through the portal may be hospital staff or outside services, and the purchase terms let Ceribell sublicense and transfer customer data for its business purposes. No subprocessor list is published.

Prospective multicenter evidence on clinical decision-making, retrospective multicenter evidence on outcomes, and algorithm validation against experts. The DECIDE study, a prospective multicenter trial of rapid response EEG's effect on clinical decisions, was published in Critical Care Medicine. SAFER-EEG, a retrospective study of 1,074 adults at four academic hospitals (2018 to 2022), found a 4.1-day shorter median ICU stay, more patients with better neurological outcomes and a door-to-EEG time 19.4 hours faster with Ceribell than with conventional EEG among matched patients, and a sub-analysis found Clarity's seizure burden associated with disability at discharge. The company's retrospective validation on 1,148 adult EEGs reported 20 of 21 status epilepticus cases alerted and one missed seizure among 726 recordings with no detection, and the pediatric and neonatal clearances rest on more than 1,700 and more than 700 patients.

The outcome comparisons are observational, so they show association rather than proof that faster detection changed outcomes.

The customer terms disclose a broad data license, and nothing states how patient recordings are protected within it. Under the purchase terms, the hospital keeps ownership of its data and grants Ceribell a worldwide, irrevocable, transferable license to use, process, store and disclose it (but not sell it) to perform the agreement and, in a form that does not identify the hospital as the source, for Ceribell's business purposes, including developing and improving its products. That is continuous EEG from critically ill adults, children and preterm neonates feeding product development under a condition that removes the hospital's identity, not necessarily the patient's.

No retention period for recordings, patient-level de-identification method, or opt-out from development use is published, and the website privacy policy's retention language is general. Safety engineering for the algorithms is not described.

A substantive customer contract reaches the service, and business associate status is only implied. The published purchase terms govern the hospital's data (ownership stays with the hospital, with a license to Ceribell) and say that any business associate agreement the parties sign is incorporated as an exhibit, which acknowledges the relationship without stating Ceribell's role, publishing the agreement or saying at which tier it is offered. The website privacy policy covers site visitors and general personal information.

Neurologists reviewing recordings remotely may not be the hospital's employees, and the system now records preterm neonates, whose records carry additional state protections; neither point is addressed.

An outside check exists through the regulator, and no attestation. The annual report (Item 1C) says the company uses NIST CSF, HIPAA, SOC 2 Type II and SOX IT general controls as guides and states that this does not imply it meets any of them; it describes risk assessments, a security team, external providers where appropriate, training, an incident response plan and vendor risk management. A security notice invites vulnerability reports, with good faith safe harbor. Its software clearances since 2023 required cybersecurity documentation reviewed by FDA as a condition of clearance, which is an outside review of product security design rather than of the company's controls.

No SOC 2, HITRUST or ISO/IEC 27001 attestation is held or named.

The clearances can be looked up in FDA's database, with their indications and dates. Ceribell's AI functions are cleared through 510(k): the Status Epilepticus Monitor (K223504, May 2023), Seizure Detection Software for patients aged one and older (K241589, April 2025), Infant Seizure Detection Software extending to preterm neonates (K252070, November 2025) and the Delirium Monitor System (K251936, December 2025), with further software clearances in 2026 (Neurology Review Software K260998, June; EEG Analysis Software K261712, September). The product page cites the clearance letters, and the company reports Breakthrough Device designation and a Medicare New Technology Add-on Payment for delirium monitoring.

The Clarity page says the algorithm keeps improving with more training cases, without saying which cleared version is deployed or how updates are handled under the clearances.

Population-specific validation exists for age, and no governance framework or evaluation across other groups is published by the company. Seizure detection was cleared separately for patients aged one and older (April 2025) and for preterm neonates from 25 weeks (November 2025), each on its own validation data (more than 1,700 pediatric and more than 700 neonatal patients, by the company's account), rather than extending adult results downward. The Clarity page states that the algorithm is tuned not to miss status epilepticus and may over-call highly abnormal activity, and that some disagreement with expert readers should be expected because experts disagree with each other.

Per-age performance figures are not on the company's pages, and nothing addresses performance by sex, hair type (which affects electrode contact for a headband) or underlying condition. The Clarity page says the algorithm is continuously improving with more training cases, and no process for reviewing or monitoring those changes is published.

An indemnity running toward the customer is published in the customer terms. Ceribell's purchase terms (last revised 23 January 2025), which govern purchase of and subscription to its products, commit Ceribell to defend, indemnify and hold the customer harmless against claims, losses and costs arising from any defect in the Product and from any claim or finding that the Product is not as represented or warranted, alongside hardware warranties and a general liability cap. The products include the Clarity software, so a hospital harmed by a defective output has a contractual route against the vendor rather than only a regulatory one.

Error rates are also published with their denominators: in the company's retrospective assessment of 1,148 adult EEGs, Clarity alerted in 20 of 21 cases of status epilepticus, and among 726 recordings where it detected no seizures, experts found seizures in one. The liability cap may limit what the indemnity is worth in practice.

Integration is claimed with no system named. The EEG portal page offers EMR integration that connects Ceribell data directly to the hospital's chart to reduce duplicate entry and speed billing, and smart EEG reports with standardized, reimbursement-ready language. No record system, interface standard, result format or live site is named.

Medication annotation happens on the device at the bedside rather than flowing from the administration record, so during a seizure a clinician records the drug twice, and the two records can diverge for whoever reads the EEG later.

A single cloud service with location implied rather than committed. Clarity is described as a cloud-based algorithm, and recordings are available to neurologists through the web and mobile portal in real time from anywhere. No hosting provider, region, residency commitment or tenancy model is published on the company's pages or in its annual report, and the recordings leave the hospital by design.

No price is published and the posture is discoverable. The annual report says revenue comes from single-patient headbands and a monthly subscription fee for Clarity, recorders and the portal; the published purchase terms describe purchased and loaned recorders (with a $2,500 charge for a lost or damaged loaned recorder), subscription terms and headband warranties. CMS granted a New Technology Add-on Payment for the delirium monitor, worth up to $2,171 per eligible Medicare inpatient case from 1 October 2026, which is payer context rather than a price.

No headband price, subscription rate or recorder price is published.

Settings and populations are named, and each is supported by its own clearance. The system is for acute care: emergency departments, adult ICUs, pediatric ICUs and neonatal ICUs, where conventional EEG and neurology reading are not immediately available. Seizure detection is cleared for adults, for patients aged one and older, and for preterm neonates from 25 weeks, each on its own validation data, and delirium monitoring has its own clearance. Outpatient neurology and epilepsy monitoring units are not the target.