Aidoc

The models are the engine of the triage modules, and the platform sold around them carries value of its own. Aidoc's cleared triage products run on CARE, the company's self developed foundation model, and do not exist without it: the January 2026 clearance for comprehensive body CT triage is a CARE product. What a health system buys, though, is aiOS, which the company presents as an enterprise platform for running clinical AI at scale.

It hosts vetted third party algorithms from partner developers and equipment manufacturers alongside Aidoc's own, and it carries care coordination (connecting care teams across workflows when a finding is flagged), patient management (identifying, capturing and following patients with findings) and systems integration services. Remove Aidoc's models and a platform running partner algorithms and a follow up workflow remains: the model is the engine of a core module rather than the whole of the product.

Triage and notification with the radiologist still reading every study; what happens after a miss is not published. Aidoc flags suspected acute findings, reorders the worklist and pushes a notification with a preview image to a desktop widget; results can also go to the native worklist, the reporting system or PACS as a separate series.

The K252970 indication states the limits: analysis runs in parallel with the standard read, preview images are for notification only, the device is not for diagnosis, and notified clinicians must view the full images. The 510(k) summary also publishes the thresholds for that product: a default balanced operating point and four alternatives per finding trading sensitivity against specificity, each with its results.

For stroke, pulmonary embolism and aortic disease, care coordination sends real time notifications with clinical context from the record, so a flag can reach specialists as well as the radiologist. The missing part is the path after the system is wrong. A false negative leaves no trace on a worklist that looks normal, and nothing published says how the platform's continuous monitoring finds misses or what a site is told when one is found. The company reports about a tenfold reduction in false alerts against unnamed single condition tools, not independently replicated.

The model is named and versioned; what it was trained on and how updates reach customers are not described. CARE (Clinical AI Reasoning Engine) is Aidoc's self developed foundation model. CARE1 was launched in December 2024 as a foundation model for CT, a successor has been announced, the company says all its models will move onto CARE, and a multi year collaboration with Amazon Web Services supports the compute.

The architectural case is explained against single task models: one model trained across conditions and anatomies that powers many indications. Training data is described only by scale and type, millions of exams of real world multimodal data, with no sites, populations, scanner mix or time period given. No update practice is published: nothing says how a customer learns that the model behind a cleared module has changed, or how a new CARE version relates to each clearance.

A short chain for Aidoc's own models, with the partner and hosting layers named and no subprocessor list. Aidoc builds its models in house on CARE, its self-developed foundation model, so no outside model provider sits between a study and a flag for Aidoc's own modules. aiOS also runs third party algorithms from partner developers and equipment manufacturers, listed on a partners page and announced by name (Cercare Medical for MR perfusion in December 2025, for example), each described as cleared or marked in its own right.

The current Security and Privacy page names AWS and Azure for the platform, and the company's documentation describes a de-identified object carrying a new identifier as what leaves the customer's environment. A partners page is not a subprocessor list: it carries no change notice, no retention or training terms for data once it reaches a partner model, and no statement of which company answers for a partner's finding. That is the gap between B and A.

The headline claim is a safety net: comprehensive body CT triage that surfaces acute findings earlier in crowded emergency departments and outpatient backlogs. The evidence reaches the accuracy half of that claim, not the in-use half.

The company's 510(k) summary for K252970 reports a retrospective, blinded, multicenter standalone study across 6 US sites, 280 cases per finding and 1,769 unique scans, judged against three senior board certified radiologists: sensitivity 94.0 to 99.3 percent and specificity 95.7 to 99.3 percent across 11 findings, every finding clearing its prespecified goal (area under the curve above 0.95). Mean time to notification was 45 seconds on 1,523 true positives.

The announcement's means (97 and 98 percent) match that table; its 'up to' figures (98.5 and 99.7 percent) do not correspond to any figure in it. A claimed tenfold reduction in false alerts against unnamed single condition tools carries no method. The research library lists journal and conference work by condition, much of it described as retrospective, and says some concerns products not yet cleared.

Homepage figures (31 percent faster pulmonary embolism notification, a 34 percent cut in stroke door to puncture time) carry reference numbers with no source shown. A white paper with Mercy, commissioned by AVIA, counts more than 50,000 new findings over five months. No prospective or in-use evaluation of the comprehensive product is published.

De-identification and retention are specific; training use is not stated. The company's security and privacy documentation describes a new DICOM object created before upload carrying only non PHI data and a newly generated identifier; analysis runs on that object, and the identifier re-attaches PHI locally when results return.

Logging and analytics data are de-identified before leaving the site, PHI is encrypted at rest in the local orchestrator, the cloud tier keeps no data beyond processing, and audit logs of data access are retained for three years. What is not stated is whether customer studies are used to train or improve models, a live question now that CARE is described as trained on millions of real world exams.

The privacy policy covers website visitors and business contacts only, and says customer information is processed on the customer's behalf and falls outside it. The detailed document dates from June 2020, before CARE. Safety engineering, meaning how a missed or false finding is handled as a safety event, is not published. There is no word on whether anything has changed since 2020.

The architecture limits what crosses the boundary; the posture itself is not stated. Aidoc's design deidentifies studies before upload and reattaches identifiers inside the customer's environment, and the company describes its solution as designed to comply with HIPAA and HITECH. That is a substantive document reaching the service.

What the company does not state is business associate status: its privacy policy says it processes customer information on the customer's behalf and excludes that information from the policy, and no agreement, tier or subcontractor term is published. The orchestrator inside the customer's environment handles PHI, so the product does touch PHI. The trust center may hold more.

ISO 27001 is claimed as the company's own, and the company runs a trust center. The Security and Privacy page links a SafeBase trust center and says the company has adopted the NIST Cybersecurity Framework, runs on AWS and Azure, and uses endpoint detection, encryption, security event monitoring and cloud posture management under a CISO led security team. The standards it says it complies with are shown as a row of logos rather than named in text.

The company's security documentation claims its own ISO 27001 certification with an annually certified information security management system, and describes quarterly vulnerability scanning of cloud resources and quarterly penetration testing against OWASP principles, with a further test on major changes. No SOC 2 report of its own is claimed; the cloud providers' SOC 2 attestations cover the infrastructure, not Aidoc. The named ISO 27001 certification is the strongest evidence here. Whether the certificate and reports can be downloaded from the trust center is not stated.

The regulatory position is stated product by product and can be checked. K252970, BriefCase-Triage: CARE Multi-triage CT Body, was cleared on 7 January 2026 under 21 CFR 892.2080 (radiological computer aided triage and notification, Class II, product codes QAS and QFM).

It covers 11 findings on contrast and non-contrast CT of the chest, abdomen and pelvis in adults: diverticulitis, abdominal or pelvic abscess, appendicitis, intestinal ischemia or pneumatosis, obstructive renal stone, small and large bowel obstruction, spleen, liver and kidney injury, and pelvic fracture. The company combines these with three findings cleared earlier to sell a 14 finding workflow.

The indication is triage in parallel with the standard read: preview images are for notification only, the device does not alter images and is not for diagnosis, and notified clinicians must view the full images. For the rest of the portfolio the company points buyers to its 510(k) notifications on the FDA database and each product's user guide, and its research library says some listed studies concern products still in development or investigational use.

The multi-triage device was announced as under review and not for sale in the US until cleared. Not cleared as a device: care coordination and patient management, and partner algorithms on aiOS, which carry their own clearances or markings. Automated draft reporting is described as in development. Match each module in a quote to its own clearance number and indication.

Governance is described as a platform function with named material behind it, and no results by patient group are published. aiOS is presented as sustaining AI performance through continuous performance monitoring and built in governance for every model it runs, including partner algorithms. The company also publishes the BRIDGE guidelines, a framework for integrating AI into clinical practice, and AI strategy and governance material written for health system leaders.

Nothing published evaluates bias or fairness with a stated method, or compares performance across the groups the product serves. The K252970 summary describes its test set by scanner manufacturer (Canon, GE, Philips, Siemens) and slice thickness and lists the other pathologies present in negative cases, but reports results only by finding, not by group.

The question is live for imaging triage: models have been shown to infer self reported race from radiographs and CT when clinicians cannot, and chest radiograph classifiers trained on public datasets have underdiagnosed women, Black and Hispanic patients, younger patients and patients on Medicaid.

An error rate with its method and denominator is published for the lead product, and nothing else stands behind the output. Aidoc's own 510(k) summary for K252970 (BriefCase-Triage: CARE Multi-triage CT Body, cleared 7 January 2026), which its announcement cites, reports a retrospective, blinded, multicenter standalone study across 6 US sites: 280 cases per indication (139 to 154 positive each, 1,769 unique scans), against a reference standard set by three senior board certified radiologists, with test cases distinct in time or site from training data.

Sensitivity runs from 94.0 percent (obstructive renal stone) to 99.3 percent (spleen injury) and specificity from 95.7 to 99.3 percent across the 11 findings, at the default operating point, with four alternative operating points reported. That is a test set result, not performance in use, and it covers those 11 findings only: not the three earlier indications in the 14 finding workflow, the neuro and vascular modules, care coordination, or partner algorithms on aiOS.

For a triage product a missed flag leaves no artifact on a worklist that looks normal, so misses in use are hard to see. The company publishes no indemnity, performance guarantee or remediation commitment. Reporting to the FDA and the public MedWatch route exist as a regulatory floor, ending in labeling changes or recalls rather than a remedy.

Deep in the imaging workflow, with the record reached for context. Aidoc ingests DICOM routed or queried from PACS and VNA or sent from modalities, plus HL7 messages, and returns results to a desktop widget, the native worklist, the reporting system, or PACS as a separate series. Authentication runs through Active Directory, a dedicated single sign on, or sign on through PACS, RIS or dictation software, a practical detail in a radiology department.

Neurovascular care coordination draws clinical context from the electronic record, and the company offers systems integration services, but no electronic record integration is named by vendor, interface or marketplace listing. The standards support has named deployments behind it (WellSpan, Mercy, Yale New Haven and others). Nothing published says which record systems are live for findings, follow up tasks or incidental finding tracking, in which direction, or through what interface.

Hybrid by design, with the processing path and tenant isolation documented and residency left open. The company's security and privacy documentation describes an orchestrator on a virtual machine inside the customer's environment that handles ingestion, routing, deidentification and reidentification, with analysis in a cloud environment. Each customer is assigned its own storage bucket with no shared production storage, and the cloud tier keeps no data beyond processing.

The current Security and Privacy page says the platform runs on AWS and Azure; the older documentation also named Google Cloud. Where the analysis runs geographically, and whether a customer can require a region, is not stated. The detailed document is revision 1.1 of June 2020, scoped to software version 2.0 and later, and predates aiOS in its current form, the CARE foundation model and the AWS collaboration, so the split may no longer hold for the current platform.

No price is published, and how the product is bought can be established before contact. Aidoc sells aiOS and its cleared modules to health systems under enterprise agreements through a demo, and maintains an AWS Marketplace listing, which makes the procurement channel public; the listing states that pricing depends on the duration and terms of the contract agreed with the vendor. The route to purchase is visible and the price is not.

Third party software directories publish figures, including a per installation starting estimate; those are aggregator estimates rather than vendor disclosures and should not be carried into a budget as though the company had published them.

Coverage is named across service lines and settings, with validation behind the cleared indications rather than the whole platform. Aidoc sells triage and follow-up across neurovascular care (stroke, hemorrhage, brain aneurysm, cervical spine), vascular care (pulmonary embolism, venous thromboembolism, aortic disease), cardiology (coronary calcium and echocardiography) and body CT, and the January 2026 clearance frames the body CT product for crowded emergency departments and for outpatient backlogs where unexpected critical findings wait.

The company reports use in more than 1,600 medical centers and names health systems including WellSpan, Mercy, Yale New Haven, St. Luke's and Renown. Validation sits at the level of each cleared indication, with its own FDA review, and the research library is organized by condition. The broader claims, that the solutions cover 75 percent of the patients in a health system and that cardiology uses text based as well as imaging AI, carry no validation by setting. The company also sells in Europe through a separate site, where the indication list may differ from the US clearances.