Optellum

The score is the product, and the regulator cleared the score rather than the software around it.

Lung Cancer Prediction AI computes a malignancy risk figure from full patterns of three dimensional pixels in a standard computed tomography scan, using a convolutional neural network functioning as a radiomics based digital biomarker. Nothing about that is separable from the model: the input is an ordinary scan already acquired for other reasons, and the entire value added is the inference performed on it. Remove the network and there is no biomarker, no score and no product.

The surrounding Virtual Nodule Clinic provides identification, triage and tracking of nodule patients, which is workflow, but it exists to route patients toward and away from the score rather than to deliver independent value. A hospital does not buy nodule tracking from this vendor and decline the prediction.

The direction of travel reinforces the grade. A thorax computed tomography foundation model announced in August 2025 and a wider platform built on it represent investment in deeper modelling rather than in broader software, which is the choice a company makes when the models are what it sells.

Graded A. A payment code describing quantitative tissue characterisation performed separately from the scan is a formal acknowledgement by a payer that the inference itself is the billable service.

The role is defined more precisely than most vendors manage, and the operating point behind it is not published.

The positioning is exact rather than general. The product is described as decision support offering second reader assurance to radiologists, and as supporting guideline aligned management decisions. Second reader is a specific and well understood role in imaging: the algorithm does not replace the primary read, it provides a check against it, and the clinician retains the decision. Framing the output as supporting guideline aligned decisions is equally precise, because it positions the score as an input to an existing published algorithm rather than as an independent recommendation.

The regulatory pathway reinforces it, since clearance was granted specifically for decision support rather than for autonomous detection.

What is missing is the threshold. No sensitivity, specificity or operating point appears in customer facing material, so a hospital cannot state what proportion of malignancies the score will fail to elevate at the cut off it deploys. Performance data exists in the regulatory submission and in the published literature, and is not gathered where a deploying clinician would find it.

One risk specific to a risk score is unaddressed. A low prediction on a nodule that is in fact malignant may reduce a clinician's inclination to pursue follow up, and the failure this product exists to fix is precisely patients lost to follow up. Nothing describes how a low score is presented, or what safeguards prevent it from becoming a reason to stop looking.

Ask for sensitivity and specificity at the deployed threshold, and how low scores are presented to clinicians.

The method is described concretely and documented in peer reviewed literature, with deployed thresholds still unpublished.

The description is specific rather than promotional. The score is computed from full patterns of three dimensional pixels in standard computed tomography images, using a convolutional neural network functioning as a radiomics based digital biomarker. Naming the architecture family, the input representation and the biomarker concept lets a technical reader understand what class of system this is, and the peer reviewed publications document the method in enough detail for independent scrutiny.

The input choice is itself a meaningful disclosure. Operating on standard scans already acquired in every modern hospital means no special protocol, no additional radiation and no new equipment, which is a practical claim a buyer can verify.

The foundation model announced in August 2025 is named and characterised as purpose built for thoracic imaging, though no technical description accompanies it.

What is missing is the operating point. No sensitivity, specificity, threshold or calibration curve appears in customer facing material, and for a continuous risk score the threshold is the entire clinical question, since the score only becomes an action when compared against a cut off. No model card exists, and the development data composition is not characterised outside the literature.

Graded B because the method is genuinely documented and externally published, and rather than higher because the numbers a deploying clinician needs are not assembled anywhere accessible.

Ask for the performance summary at deployed thresholds and the development population composition.

The models are built rather than assembled, and the chain around them is undisclosed.

The substantive point in the vendor's favour is that the intelligence is proprietary and developed in house. A radiomics based convolutional network developed as a digital biomarker, documented in peer reviewed literature and cleared by four regulators, is not a third party model with a wrapper, and the thorax computed tomography foundation model announced in August 2025 is described as built by the company. That removes the largest supply chain question this index usually asks, since no external model provider sits in the inference path.

The information security certification implies documented supplier management, since that standard requires it, though no supplier is named.

What remains undisclosed is everything else. No cloud or hosting provider, no sub processor register, no machine learning framework or third party component, and no statement on whether customer imaging contributes to model development.

Training provenance is the more consequential gap and is partly resolvable from the literature rather than from the company. Development and validation datasets for this class of model come from named cohorts and institutional collections, frequently including national screening trial data, and the governance under which those were obtained matters for a commercial product built on them. Nothing gathers that information in one place for a buyer.

The foundation model sharpens the question, since foundation scale training implies a substantially larger corpus than the original biomarker required, and its origin is unstated.

Ask for the development and foundation model dataset provenance and governance, the hosting provider, and whether customer scans feed model updates.

Evidence that runs the full distance from algorithm validation to health economics, which almost nothing else in this index does.

The foundation is external validation published independently of the company's own marketing. A convolutional neural network validation appeared in Thorax in 2020, and an assessment of deep learning risk stratification for indeterminate pulmonary nodules appeared in the American Journal of Respiratory and Critical Care Medicine the same year. External validation, meaning performance measured on data the model was not developed against, is the strongest form available short of a trial.

Real world impact followed. Findings published in BMJ Open Respiratory Research in December 2025 and presented at a European oncology congress reported accelerated time to diagnosis and improved follow up rates. Follow up rate is the right outcome for this problem, because the clinical failure it addresses is not misreading a scan but losing the patient afterwards, and a large share of incidental nodule patients never return.

The economic analysis is the rarest element. In March 2026 a Vanderbilt led lifetime payer perspective study reported the approach cost effective, and a lifetime horizon payer analysis is a category of evidence this index has not encountered from any other vendor. Competitive national research funding awarded in June 2026 for a multi site health service pilot adds a further independent assessment.

Deployment corroborates scale: more than 250 clinical sites and over three million scans analysed by June 2026, with named academic customers including two major university health systems and a large teaching hospital.

The qualification is authorship and sponsorship, which are not always separable from the company across this body of work.

Ask for the BMJ analysis design and comparator, and the cost effectiveness assumptions.

A certified management system with an unusually explicit scope, and little published detail about handling itself.

The certification carries more information than most because the scope is enumerated rather than implied. The information security management system is stated to cover design, development, manufacture, sales and support of clinical decision support software for lung disease, which means the certified boundary includes the development pipeline where models are built and the support function where customer data is most likely to be touched. Many certifications cover corporate infrastructure and quietly exclude the product; this one does not.

A public trust centre provides a route to the underlying documentation, which is the right mechanism even when the contents are gated.

What is absent from open material is the operational detail this index looks for: no encryption statement, no retention schedule, no access control description, no data location statement and no deletion process was located publicly.

The longitudinal design makes one of those materially more important than it would otherwise be. Tracking a nodule over time requires retaining prior imaging and prior scores, so the system accumulates a patient specific cancer risk history. A stored record that a named patient carried an elevated malignancy prediction is sensitive in ways an ordinary scan is not, with implications for insurance and employment that outlast the clinical episode.

Whether customer imaging contributes to model development is also unstated, and the foundation model announced in 2025 makes that question more pointed rather than less.

Ask what is retained for longitudinal tracking, where it sits, and whether customer scans train models.

A trust centre exists as the route to compliance documentation, and the specific health privacy position is not publicly readable.

The supporting structure is real. A public trust centre runs on an established compliance platform of the type that ordinarily carries agreement templates, subprocessor lists and compliance attestations behind an access request. Certification against the 2022 revision of the international information security management standard was achieved in August 2024, with the scope enumerated across design, development, manufacture, sales and support of clinical decision support software.

United States exposure is unambiguous and substantial. The product is deployed at academic and community health systems, billed to the federal payer under assigned procedure codes, and processes identifiable imaging in the ordinary course, so business associate agreements necessarily exist across the installed base.

What could not be established publicly is the posture itself: no agreement template, execution requirement, negotiation stance or subcontractor flow down position is readable without requesting access, and no explicit statement addressing federal health privacy law was located in open material.

One question specific to this product remains unanswered anywhere. Nodule tracking is longitudinal by design, comparing scans across months and years, which means the system holds a patient imaging history rather than processing a single study and discarding it. Retention period, storage location and deletion on request are the natural follow ups and none is public.

Ask for the agreement template through the trust centre, the retention period for longitudinal nodule tracking, and the deletion process.

A public trust centre alongside a certification named to the revision, with the certified scope stated explicitly.

The certification disclosure is the most precise recorded in this session. The standard is named, the 2022 revision is specified, the achievement date of August 2024 is given, and the scope of the management system is enumerated across design, development, manufacture, sales and support of clinical decision support software for lung disease. Each of those elements does work. The revision distinguishes the current control set from its predecessor. The scope statement establishes that the certified boundary includes the product development pipeline rather than only corporate systems, which is the distinction that determines whether a certificate means anything for the software a hospital is buying. This index has repeatedly docked vendors for naming a standard by family without the number, the revision or the scope; this vendor supplies all three.

A public trust centre runs on an established compliance platform, providing a standing route to security documentation without a sales conversation.

Three gaps remain. No service organisation controls report is claimed, which United States health system procurement increasingly expects and which matters given substantial United States deployment. No certificate number, certification body or audit period is published in open material. And no penetration testing statement or vulnerability disclosure policy was located.

Graded A on the strength of a public trust centre and a certification disclosed with genuine precision. Ask for the certificate number and body, whether a controls report exists for the United States market, and the vulnerability disclosure route.

Clearance in four jurisdictions plus an assigned payment classification, which together represent the most complete regulatory position in this index.

United States clearance was obtained in March 2021 and was the first granted for artificial intelligence decision support in lung cancer diagnosis, which means the company defined the regulatory path rather than following one. European medical device regulation marking and United Kingdom marking followed, and Australian class IIb approval was granted in February 2026 after review of technical documentation and validation studies. Four separate regulators examined this software and permitted it.

The payment classification is a distinct achievement and belongs on this axis rather than only on the commercial one. Securing procedure codes and a payment classification requires demonstrating to a payer that a service is clinically distinct and resource consuming enough to warrant separate payment, which is a substantive review conducted by a different body against different criteria. Very few artificial intelligence products in this index have cleared both a regulator and a payer.

The company is also honest about what it holds. Reimbursement material states plainly that the codes are temporary, carry no assigned relative value units and may be declined case by case by regional contractors. That is an accurate description of an early payment pathway rather than an overstatement of it, and this index has recorded several vendors doing the opposite with weaker credentials.

What is not enumerated publicly is the indication for use as cleared, or the performance data submitted.

Ask for the clearance number and indication for use, and the current status of the temporary codes.

A substantial external validation record, and no performance stratified by any subgroup or by scanner.

What supports the grade is the validation base rather than a governance framework. Two independently published external validation studies, real world evaluation in a third journal and an economic analysis led from an academic centre together constitute more third party examination than most vendors here have received. External validation on data the model did not see is itself partially protective against the narrowest forms of overfitting.

What is absent is stratification. No performance broken down by sex, age, race or smoking history was located, no fairness testing, no calibration analysis by subgroup and no external algorithmic audit.

Two bias axes are specific to this product. The technical one is scanner variation: computed tomography output differs by manufacturer, model, slice thickness and reconstruction kernel, and radiomics features are known to be sensitive to exactly those parameters, more so than coarser detection tasks. Because newer scanners concentrate in better resourced institutions, accuracy that tracks acquisition quality tracks institutional wealth.

The clinical one concerns the population. Nodule malignancy prediction depends on features whose distribution varies with smoking history, sex and ethnicity, and lung cancer incidence and presentation differ across those groups. A model calibrated on one population can be systematically miscalibrated in another, and calibration rather than discrimination is what matters when a score drives a follow up decision.

Ask for calibration by sex, race and smoking history, and performance across scanner manufacturers and reconstruction protocols.

No contractual allocation is published, in a product whose output can reasonably influence a decision not to investigate a cancer.

A dedicated pass located no service level agreement, no accuracy warranty, no uptime commitment, no indemnity and no remediation position.

The mitigating structure is genuine and belongs on the record without moving the grade. Clearance in four jurisdictions brings post market surveillance obligations, adverse event reporting and change control on the algorithm, so material modification cannot be made silently and there is a regulator to escalate to. That is more continuing accountability than any unregulated competitor carries.

What that does not address is the specific failure mode. This is a risk score, and the clinical harm from a score that is too low is not a missed detection on a single read but a patient whose follow up is deprioritised and whose cancer is found later. The product exists because follow up failure is endemic, so a tool that occasionally provides false reassurance is operating on exactly the failure it was bought to prevent. No published miss rate at the deployed threshold means that residual risk cannot be quantified by the hospital accepting it.

The reimbursement pathway adds a second exposure that is unusual in this index. A hospital billing a payer for a service performed by this software is exposed to claim denial and to documentation requirements set by regional contractors, and nothing states what support or indemnity the vendor provides if reimbursement is refused.

One pre emptive note: further clearances or publications cannot move this grade. Only contractual terms or a published operating point will.

Ask what is warranted on performance, and what support exists if claims are denied.

Integration with both the record system and the imaging archive is claimed, and neither is specified.

The claim is broader than most vendors in this category make, since imaging artificial intelligence typically integrates with the picture archiving system alone and this product asserts both. That breadth is plausible given the workflow: nodule tracking requires patient level context and follow up scheduling that lives in the record system rather than in the imaging archive, so a purely imaging integration would not support the product's core function.

What is absent is every specific. No electronic health record is named, no imaging archive vendor is named, no interface standard such as DICOM, HL7 or FHIR is described, no marketplace or validated integration listing was located, and no statement addresses which direction data flows or in what format results are returned.

The gap is more consequential here than for a pure detection tool. A risk score that does not enter the record cannot drive a follow up reminder, and follow up failure is the specific clinical problem this product exists to solve. The published evidence reports improved follow up rates, which implies the integration works, and public material does not explain how.

One related question is unanswered. The billing pathway requires the service to be documented and coded, which means the result must reach the systems used to raise a claim, and nothing describes that path.

Ask which record and imaging systems are integrated in production, through what standards, whether the score posts as a discrete result, and how the billable service is documented.

A certified scope and a trust centre imply the information exists, and none of it is publicly readable.

A dedicated pass located no hosting provider, no deployment options, no region statement, no residency commitment, no tenancy description and no statement of whether imaging is processed locally or transmitted for analysis.

What can be established indirectly is modest but real. The information security management certification covers design, development, manufacture, sales and support of the software, which indicates a documented environment behind the product, and a public trust centre exists as the route through which deployment and hosting documentation would ordinarily be requested. That is a disclosure mechanism, gated rather than absent.

The unanswered question matters because of the geography. The product is approved and deployed across the United States, Europe, the United Kingdom and Australia, and European deployment carries data location and transfer requirements that a global imaging service must satisfy. Nothing public describes whether processing is regionalised, whether scans leave the hospital at all, or whether an on premise option exists.

The comparison within this lane is instructive. Another imaging vendor indexed here states in open material that patient data remains on a server inside the hospital and is not transmitted externally, which settles the question before a buyer asks. This vendor may well have an equally good answer and has not published it.

Ask whether processing occurs on premise or remotely, where hosted processing takes place, and how European deployments are handled.

The revenue side is published with figures and honest caveats, and the cost side is not published at all.

What exists is the most substantive commercial disclosure recorded in this session. A dedicated reimbursement page sets out two temporary procedure codes effective 1 July 2022 covering quantitative computed tomography tissue characterisation, one performed standalone and one alongside a concurrent scan. In the hospital outpatient setting the standalone code carries a new technology payment classification at roughly 600 to 700 dollars per use, paid at the full rate without multiple procedure discounting. A hospital can therefore calculate revenue per use before speaking to anyone.

The candour matters as much as the figure. The company states plainly that these are temporary category codes, that they carry no assigned relative value units, that claims may require supporting documentation to establish medical necessity, and that regional contractors decide reimbursement case by case. Publishing the weaknesses of your own payment pathway is disclosure against interest, and this index grades that behaviour rather than the pathway.

What is missing is what the hospital pays. No licence fee, unit of charge, per site cost, per scan cost, implementation fee or contract term was located. Reimbursement is not price, and a buyer with a precise revenue figure and no cost figure still cannot compute a margin.

Graded B because half the commercial equation is published with unusual rigour and the half the vendor controls is absent. Ask for the licence basis and rate, and how it relates to expected reimbursement volume.

One disease, addressed across the whole pathway and four regulatory jurisdictions.

The clinical scope is deliberately singular: pulmonary nodules and early lung cancer. Within it the coverage is genuinely end to end, spanning identification of at risk patients, triage and prioritisation, risk stratification of indeterminate nodules, tracking over time, and second reader assurance for radiologists reading nodules on computed tomography. The product addresses both the incidental nodule found on a scan ordered for something else and the screening context, which are different clinical populations with different prior probabilities.

Geographic coverage is evidenced by regulatory approvals rather than asserted, spanning the United States, Europe, the United Kingdom and Australia, with the Australian approval obtained in February 2026 after review of technical documentation and validation studies. Named deployments cover United States academic health systems, a United Kingdom teaching hospital and community health systems, so the installed base is not confined to research centres.

What is not evidenced is the mix. More than 250 sites are claimed and nothing states how many run the full nodule programme rather than the score alone, or how deployment divides between radiology and pulmonology, which are different buyers with different workflows.

A collaboration announced in June 2026 extends into clinical trial matching for non small cell lung cancer, which is an adjacent setting rather than a new disease.

Ask how the installed base divides between radiology and pulmonology, and between screening and incidental nodule programmes.