Optellum

The model is the engine of a core module, and the product around it carries its own value. Optellum's Lung Cancer Prediction AI computes a malignancy score for a pulmonary nodule from the CT scan with a convolutional neural network (LCP-CNN in the 510(k) summary, K202300), and that score does not exist without the model.

It is sold inside the Virtual Nodule Clinic, which the company describes as its early lung cancer diagnosis solution and which also identifies patients with nodules, triages them, and tracks and manages their follow up; the 510(k) summary gives nurses, clinicians and navigators access to the management functions alone.

Tracking patients who would otherwise be lost to follow up is value a hospital gets whatever the score says, and the company's published impact claims are about that pathway. The company also describes LungOS, a wider platform on a thorax CT foundation model announced in 2025.

The prediction is decision support with its users and scope defined, and the operating point is not published. The 510(k) summary (K202300, March 2021) indicates the Lung Cancer Prediction score for characterizing incidentally detected solid and semi solid nodules of 5 to 30 mm in patients aged 35 and over, as an adjunct to decisions made under published guidelines; pulmonologists and radiologists see the score, while nurses, clinicians and navigators use only the management functions.

The network's output, a value between 0 and 1, is mapped to an integer from 1 (most likely benign) to 10 (most likely malignant) and shown against a histogram from a reference database, so a clinician sees where a nodule sits rather than a single verdict. The company positions it as a second reader for radiologists.

What is missing is the threshold: no sensitivity or specificity at any score is published, and nothing says how a low score is presented or what stops it from becoming a reason to stop following a nodule, which is the failure the product exists to prevent.

The method and the training data are described, and versioning is not. The 510(k) summary (K202300) names the model, LCP-CNN, a convolutional neural network applied to a region of interest the user selects on the CT, and characterizes its training data: about 8 percent malignant and 92 percent benign nodules of 5 mm and larger, 95 percent from the United States and 5 percent from Europe, 38 percent from women, median age 62 (range 20 to 90), and 95 percent from screening with 5 percent incidentally detected.

The method is also described in peer reviewed validation papers (Thorax, 2020; American Journal of Respiratory and Critical Care Medicine, 2020). The cleared population, incidentally detected nodules from age 35, differs from the mostly screening training mix. The company announced a thorax CT foundation model in August 2025 with no technical description. No model version or update practice is published, so a buyer cannot tell when the deployed model changes.

The chain is structurally short. The prediction model is Optellum's own, cleared by the FDA (K202300) and marked in the UK and Europe, so it cannot be swapped for another supplier's model without going back through those regulators, and no outside model provider sits in the inference path. The 510(k) summary describes the web application on a virtual server in the hospital's data center and the network module on a GPU server on the hospital's premises or in the cloud.

The trust center lists certifications and no subprocessors; no cloud or hosting provider is named for the cloud option, and whether customer scans contribute to model development, including the foundation model announced in 2025, is not stated.

The evidence runs from external validation to real world studies, and the two newest point different ways on the company's headline claim. External validation of the convolutional neural network appeared in Thorax (2020), and an assessment of its risk stratification of indeterminate nodules in the American Journal of Respiratory and Critical Care Medicine (2020).

The 510(k) summary (K202300) reports a retrospective reader study of 300 cases (150 malignant, 150 benign; 174 from the US and 126 from Europe; 175 from academic and 125 from community centers) in which 12 readers (6 radiologists, 6 pulmonologists) raised their mean AUC from 81.9 to 88.8 with the score, and a standalone AUC of 0.867.

A before and after study at three US lung nodule clinics (Jupiter Medical Center, AtlantiCare and the University of Connecticut; matched 12.5 month periods; posted on medRxiv in November 2025, not yet peer reviewed, with three company employees among the authors) found new patients rising from 138 to 232 and invasive procedures from 5.0 to 7.8 a month, while median time from baseline CT to diagnosis was unchanged, 58 days before and 57 after (p = 0.28).

The study the company announced in BMJ Open Respiratory Research in December 2025, funded by Bristol Myers Squibb, is described in that announcement as a retrospective comparison, at one US academic center, of standard practice with a modeled AI enabled workflow; it reports follow up compliance rising from 34 to 94 percent and median time to diagnosis of non small cell lung cancer falling from 129 to 25 days.

The home page claims 71 percent more interventional procedures, a 50 percent reduction in time to diagnosis and a tripled chance of survival. None of the three figures matches a published study, and the one prospective measurement of time to diagnosis found no change. Taken together, this is published research that stops short of independent, prospective validation of the outcomes claimed.

A Vanderbilt led lifetime payer analysis reported the approach cost effective in March 2026, which is a model rather than a measurement. Much of this body of work is authored with or funded by the company.

The company publishes the frame of a security program and little about what happens to patient imaging. Its trust center lists Cyber Essentials and Cyber Essentials Plus, ISO/IEC 27001, ISO 13485 and the NHS Data Security and Protection Toolkit, with documents available on request, and the 510(k) summary describes deployment on hospital servers, with the network module on premises or in the cloud.

The privacy policy (last updated May 2026) says personal data will not be processed for machine learning when the company uses large language models for insights, and sets retention at up to six years after an engagement ends. Both commitments are about the people the policy covers, and the policy does not say it reaches the imaging the software processes.

The company publishes no retention period, training use statement or deidentification practice for patient data, so what it offers on patient data is general assurance. Tracking a nodule over time means the system keeps prior imaging and prior scores, a patient specific record of cancer risk that is more sensitive than an ordinary scan. Safety engineering, meaning guardrails and how a safety event is handled, is not published.

Optellum sells into US health systems, its analysis carries Medicare billing codes, and no business associate position is stated anywhere the company publishes. The privacy policy (Optellum Ltd, Oxford, last updated May 2026) cites UK data protection law and describes the company's own handling of personal data, with retention of up to six years after an engagement ends.

It does not say whether it reaches the software or the patient imaging the software processes, and it contains no HIPAA, business associate or processor statement. The trust center lists the NHS Data Security and Protection Toolkit, the assessment an organization handling NHS patient data completes, beside its security certifications, with documents on request. The result is a compliance position under the UK regime, with no document that reaches the service and no US position stated.

The 510(k) summary describes deployment inside the hospital's data center, with the scoring module on premises or in the cloud, so what the company holds depends on the option chosen, and longitudinal tracking means a patient's imaging and scores are kept over time.

Recognized certifications are named, and the artifacts sit behind a request. Optellum's trust center, on the SafeBase platform and linked from the site footer, lists ISO/IEC 27001, Cyber Essentials and Cyber Essentials Plus, ISO 13485 (a quality standard) and the NHS Data Security and Protection Toolkit, and invites buyers to request access to its security documentation.

The company has announced ISO/IEC 27001:2022 certification in August 2024, with a scope covering the design, development, manufacture, sales and support of its lung disease software; an ISO 27001 certificate runs three years with annual surveillance audits. The certificate number, certifying body and any penetration test summary are not public, and no SOC 2 report is listed, which US procurement often asks for.

Optellum's US position is identifiable and matched to the product. The Optellum Virtual Nodule Clinic (Optellum software, Optellum platform) was cleared on 5 March 2021 (K202300, 21 CFR 892.2060, product code POK, computer aided diagnosis), with the breast CADx software QuantX (DEN170022) as predicate.

The indication covers characterizing incidentally detected solid and semi solid pulmonary nodules of 5 to 30 mm in patients aged 35 and over, as an adjunct to guideline based decisions, with the malignancy score for pulmonologists and radiologists and the management functions for nurses, clinicians and navigators. The company describes the product as UKCA and CE marked, and Australian class IIb approval was reported in February 2026.

The company also sells to lung cancer screening programs, while the cleared indication names incidentally detected nodules, and it does not say how the score is used in the US on nodules found at screening.

The 510(k) summary discloses a subgroup evaluation with its groups and publishes only the conclusion. For the reader study (K202300), Optellum reports no significant differences in performance by academic or community center, screening or incidental detection, reader specialty, patient sex, age range, CT manufacturer (GE, Siemens, Philips, Canon/Toshiba), slice thickness, reconstruction algorithm or dose, and gives counts for some groups (174 US and 126 European cases; 175 academic and 125 community; 206 incidental and 94 screening; 243 solid and 57 semi solid nodules).

No per group figures are published, so a buyer cannot see where performance is weakest, and a small group may show no difference only because it is small. Two groups matter most for this product: scanner and reconstruction settings, to which radiomic features are sensitive, and populations whose smoking history, sex and ethnicity change nodule malignancy risk, where calibration rather than discrimination drives the follow up decision. Smoking history and race are not among the groups reported.

Mechanisms exist that let a clinician weigh the output, and nothing stands behind it. The score is an adjunct, shown against a reference histogram, for pulmonologists and radiologists to use with published guidelines (K202300). The 510(k) summary reports a standalone AUC of 0.867 on an independent validation set with no operating point or case count beside it, and a reader study in which readers' AUC rose with the score; neither is an error rate at the score cut offs sites act on.

No service level agreement, accuracy warranty, indemnity or remediation commitment is published. The failure that matters is a low score on a malignant nodule, which can lead to a patient's follow up being deprioritized and the cancer found later. The reimbursement pathway adds a second exposure: nothing states what support the company gives a hospital if a payer denies the claim.

Integration with the record and the imaging archive is claimed, and neither is specified on the home page. The company says the Virtual Nodule Clinic integrates with a hospital's EHR and PACS, and its footer links an interoperability page, which is where named systems and standards would be set out. The 510(k) summary describes a web application on a hospital server, with a CT region of interest chosen by the user.

Nodule tracking needs patient level context and follow up scheduling from the record, and the billing codes need the result documented, so how the score reaches the chart matters more here than for a detection tool.

Deployment options are stated, and a location for the cloud option is not. The 510(k) summary (K202300) describes the web application on a virtual server in the hospital's data center and the neural network module on a GPU server on the hospital's premises or in the cloud. The trust center lists certifications and no hosting provider or region, and no residency commitment is published for the cloud option. The product is sold in the United States, Europe, the United Kingdom and Australia, where data location and transfer rules differ.

No vendor price is published, and the payer side is published in detail. The company's reimbursement page sets out two temporary procedure codes effective 1 July 2022 for quantitative CT tissue characterization, one standalone and one with a concurrent scan; in the hospital outpatient setting the standalone code sits in a new technology payment classification at roughly $600 to $700 per use, paid without multiple procedure discounting.

The page states plainly that the codes are temporary, carry no relative value units, may need documentation of medical necessity, and are decided case by case by regional contractors. That figure is what the payer pays the hospital, not what Optellum charges. No license fee, unit of charge, implementation fee or contract term is published, so a buyer can see how the service is paid for and not what it costs.

Coverage is one disease across the pathway, with validation behind part of it. The cleared indication covers incidentally detected solid and semi solid nodules of 5 to 30 mm in patients aged 35 and over, and the 510(k) reader study drew 300 cases from US and European, academic and community sites, 206 incidental and 94 from screening. The company sells to incidental nodule and lung cancer screening programs, with approvals reported in the United States, Europe, the UK and Australia.

Named users include the University of Pennsylvania Health System, Vanderbilt University Medical Center, Oxford University Hospitals and Tanner Health, and the company reports more than 250 sites. How many run the full nodule program rather than the score alone, and how use divides between radiology and pulmonology, is not stated.