Cytovale IntelliSep

The model is the engine of the result, and the signal comes from the measurement. The Cytovale System stretches tens of thousands of white cells through a microfluidic junction and films them; a neural network segments each cell, shape and deformation features are computed, and a regularized logistic regression trained on patient data turns the features into the IntelliSep Index. Without the model there is no score.

But the discriminating information is the cell biophysics: the development paper identified two engineered measurements, aspect ratio and a visco-elastic inertial response, as more informative than cell size or count. The company presents the sepsis test as the first application of a measurement platform. That places the learned component as the engine of a core module rather than the whole product.

What the test may and may not do is published, with the thresholds and what each implies. The cleared indication is adults with signs and symptoms of infection in the emergency department, to aid early detection of sepsis with organ dysfunction within three days. The labeling states that it is not to be used as the sole basis for a sepsis diagnosis, does not identify the organism, was not evaluated for repeat testing or beyond the first emergency department encounter, and has no established performance in the listed clinical groups. The score is cut into three bands at published points (4.9 and 6.2), each with its predictive value and likelihood ratio, so Band 2, the middle one, is a stated indeterminate zone rather than a forced call. The instrument runs internal quality checks and reports an error instead of a result when they fail. Every result goes to a clinician as one input among others.

The marketing leans harder than the label: the clinicians page cites a 98 percent negative predictive value without stating its endpoint, the zero sepsis-associated deaths claimed for Band 1 comes from a pooled analysis limited to patients whose sepsis label the review panel and the site physician agreed on, and one hospital's protocol routes Band 1 patients to a localized or no infection pathway.

The model, its training data and its change history are all on the public record. The development paper (PLOS ONE, 2021) describes the pipeline: an object detection step finds the frames containing cells in about four million frames of video, a convolutional neural network segments each cell from the background, descriptors of shape and deformation are computed for each cell and summarized into features, and an L1-regularized logistic regression built with scikit-learn converts the selected features into the index. It was trained and cross-validated on a high acuity emergency cohort of 307 patients collected in 2016, with about a quarter held back as a sequestered test set, then checked on a low acuity cohort of 94 and on 72 healthy donors.

The paper used an early prototype, and any retraining for the cleared version is not described. Changes since clearance are documented in FDA summaries: the 2025 software clearance states that the result calculation is unchanged since 2022 and that a sequestered data set was reprocessed under good machine learning practice, and the closed instrument accepts no incoming data, so the model cannot change without a regulated update.

A structurally short chain. The model is Cytovale's own, cleared, and runs on a workstation inside the instrument in the hospital laboratory; the instrument accepts no incoming data, so the model cannot be swapped without a regulated update, and results leave only through the hospital's own network.

The optional log export added in 2025 carries no patient data, and its destination is not named. The workstation's manufacturer is not named either, though the 2025 clearance records that its processors were replaced because the original hardware was being discontinued. The company's security page links to a Vanta trust center that may list subprocessors, but the portal requires a login.

A prospective multicenter validation is published, and it shows moderate discrimination. The pivotal study (Critical Care Explorations, 2024) tested 572 adults with suspected infection at five emergency departments in four states, with sepsis adjudicated by Sepsis-3 criteria by physicians blinded to the result: sepsis was present in 11.1 percent of Band 1, 28.1 percent of Band 2 and 49.4 percent of Band 3, with an area under the curve of 0.74 and a median turnaround of 7.2 minutes.

A pooled analysis of 1,002 patients from seven sites (Academic Emergency Medicine, 2024) reported sensitivity above 92 percent for the low band, but it kept only patients whose sepsis label the review panel and the site physician agreed on. An independent study at the University of Kansas (Journal of Clinical Medicine, 2024) compared the index with other markers in 44 patients. One hospital's pre and post study (Healthcare, 2025) reported sepsis-associated mortality falling from 10.9 to 6.6 percent after a quality program that included the test, with Cytovale staff helping analyze the data; the design cannot separate the test from the program around it. Company authors or company assistance appear on most of these papers.

Categorical statements about where patient data goes are published, without the retention detail behind them. The analyzer is a closed benchtop instrument that sends results to the hospital's middleware and laboratory information system. The page for laboratory staff still says it does not connect to the internet; since the March 2025 clearance (K250513) an optional connection, enabled after the hospital's IT review, exports system logs for remote troubleshooting, and the clearance summary states that the export is one way, that the instrument accepts no incoming data, and that the logs carry no protected health information. Customer specimens therefore do not flow back to train the model.

What stays on the instrument is not described: how long the cell images and results persist, whether they are purged, and what a service engineer may take away. The security page mentions defined retention policies and deletion of customer data at offboarding without the schedule. Cytovale publishes no safety engineering, meaning guardrails, hallucination mitigation and how a safety event is handled; the nearest thing is the instrument's internal quality checks, which withhold a result rather than report a doubtful one.

A no-access position is implied by the design and stated in pieces, without a document that reaches the service. The laboratory page says the analyzer connects only to hospital middleware and not to the internet, and the 2025 clearance summary says the optional log export carries no protected health information; on that design the company would not receive patient data in normal operation.

But no business associate position is stated either way, the only privacy statement covers the website, and the service route, where an engineer may handle an instrument that holds patient results and images, has no published terms. The laboratory page's no-internet statement is also out of date against the 2025 log export option. Nothing published says whether the company signs a business associate agreement for service access.

A recognized attestation is named, with the report held back. The company's security page states SOC 2 Type II attestation against the security criteria and offers the full report on request through a Vanta trust center; the portal sits behind a login, so the audit period, auditor and scope are not public. The same page lists regular penetration testing, encryption across the platform, access controls and tested continuity plans.

For the instrument itself, the 2025 clearance summary records independent penetration testing of the new outbound connection and review under FDA's premarket cybersecurity guidance, and the laboratory page offers a manufacturer disclosure statement for medical device security on request. Whether the SOC 2 scope reaches the instrument and service processes, or only the corporate systems, is not stated.

Three clearances are identifiable, each with its scope. IntelliSep was cleared as a Class II device under 21 CFR 866.3215 (product code QUT) in K220991 on 20 December 2022, with a procalcitonin assay as the predicate. K250513 (21 March 2025) cleared software and hardware changes: refined internal quality checks, an additional image analysis workstation variant, and an optional outbound connection for log export; the summary states the result calculation is unchanged and that reprocessed clinical and sequestered data showed no performance difference. K250912 (19 November 2025) covered changes to the cartridge and its packaging. None changed the indication: adults with signs and symptoms of infection in the emergency department, to aid early detection of sepsis with organ dysfunction within three days, alongside clinical assessment.

The company describes the test as the first and only FDA-cleared test for sepsis detection; that is the company's claim, and it turns on the difference between clearance and other routes to market.

Subgroup performance is published, and it shows where the test is weaker. The FDA decision summary for the original clearance (K220991) reports predictive values per band by age and by sex, with 80 percent confidence intervals, from the 572 patient validation study. The share of Band 3 patients adjudicated septic was 67.9 percent at 65 and over against 39.4 percent under 65, and 59.5 percent in men against 38.2 percent in women; Band 1 held between 8.5 and 14.3 percent across all four groups. The analysis is stated to be descriptive and not powered, and race and ethnicity are reported only as the spread of patients across bands.

The labeling names the clinical groups where performance has not been established: people with hematologic cancers or marrow disorders, transplant recipients, recent cytotoxic chemotherapy, nursing facility residents, and people on systemic corticosteroids or dialysis, whom the study included but could not evaluate. Those are the populations in which white cell behavior is most likely to shift for reasons other than sepsis.

A falsifiable performance record is published, and nothing stands behind it. The FDA decision summary and the validation paper give, for each band, the number of patients adjudicated septic and not septic, the predictive value and the likelihood ratio, under three sepsis definitions and three adjudication schemes: under Sepsis-3, 28 of 252 Band 1 patients (11.1 percent) were septic, and 79 of 160 in Band 3 (49.4 percent). A buyer can hold the test to those figures in local validation.

No warranty on result accuracy, indemnity or remediation commitment is published, and no purchase terms are public.

Integration is claimed through the hospital's middleware with no system named. The product page says results are delivered automatically to the laboratory information system at the end of each run, and the laboratory page says setup needs only an ethernet connection to hospital middleware; no middleware product, laboratory system, interface standard or record system is named.

The one hospital case study describes ordering the test through Epic alerts and documenting the band in Epic notes, but that is the hospital's own configuration of its record system rather than an integration the company supplies. For a test whose value depends on an eight minute result reaching a triage decision, how the result surfaces depends on each site's own configuration.

One deployment option, stated plainly, with the storage path left open. The test runs on a benchtop analyzer in the hospital laboratory, from 100 microliters of EDTA whole blood, with a result in about eight minutes delivered to the laboratory information system over the hospital network; image analysis runs on a workstation module inside the instrument. Patient data therefore rests in the hospital.

An optional outbound connection added in 2025 exports system logs without patient data, to a destination that is not named. How long images and results persist on the instrument is not stated. The laboratory page asks for four feet of bench space and a one-week installation.

No price is published, and the cost structure is discoverable. The model is a capital instrument, the Cytovale System, plus a single-use cartridge and reagents for each test, with quality control material bought from the company or made by the laboratory. The laboratory page states the footprint and service load: four feet of bench, a one-week installation, no calibration and preventive maintenance every 90 days.

No instrument price, cartridge price or service charge is given. The company's white paper argues the case through sepsis coding and reimbursement rather than test cost; an application for a Medicare new technology add on payment appears on CMS's public register.

Deliberately narrow, and validated where it claims to work. The setting is the adult emergency department at the point of triage, for patients with signs and symptoms of infection, with a three-day prediction window. That is the population of the pivotal study: 572 adults at five academic and community emergency departments in Louisiana, Missouri, North Carolina and Washington, spanning ages 18 to 103.

The labeling states what lies outside it: children, repeat testing, use after the first emergency department encounter, and clinical groups in which performance has not been established. No inpatient, intensive care or pediatric use is claimed.