Aidoc

The detection models are the product and the company built its own foundation model to produce them. CARE is not a wrapper on third party models; it is self developed and is what the FDA reviewed.

Triage and notification, with the radiologist reading every study. Aidoc reprioritises the worklist and pushes a notification carrying the key image to a desktop widget, and distributes results to the native worklist, the reporting system or back to PACS as a separate new series in the study. Nothing is removed from human review, which places the product at the opposite end of the autonomy spectrum from the autonomous reporting products in this index. The January 2026 clearance is for triage specifically.

The live question for this design is alert burden rather than missed oversight. The company reports roughly an order of magnitude reduction in false alerts against single condition tools, which is a vendor generated comparison against unnamed products and has not been independently replicated.

Unusually specific. The foundation model is named (CARE, for Clinical AI Reasoning Engine), its architectural rationale is explained against single task models, a successor version CARE2 is publicly announced, and the compute partnership supporting it is disclosed. Model versioning is visible rather than inferred, which is exactly what this axis asks for.

Aidoc discloses the layer most platforms in this index conceal, which is whose algorithm actually produced a given result. The aiOS platform is presented as vendor neutral and runs third party algorithms alongside Aidoc's own, a partners page is published for vetted algorithm developers and original equipment manufacturers, and individual partnerships are announced by name with the capability attributed, such as Cercare Medical for magnetic resonance perfusion in the neurosciences set announced in December 2025, with partner algorithms described as cleared or marked in their own right rather than absorbed silently into the platform.

Aidoc's own model layer is named as well, the CARE foundation model with a successor publicly announced, and the compute collaboration behind it disclosed. The hosting tier is enumerated too: analysis runs on Amazon Web Services, Google Cloud Platform or Microsoft Azure, with a unique storage bucket per customer and no shared production storage.

The protected health information path is documented rather than asserted, since a de identified object carrying a newly generated identifier is what leaves the customer environment and reattachment happens locally. Graded at the top because the buyer can see the model layer, the partner layer, the hosting layer and the data path, all by name. Two residuals frame the next questions rather than reduce the grade.

A partners page is marketing collateral, not a subprocessor list: it enumerates who is in the chain but carries no change notification mechanism, so a customer is not told when the set changes, which is the one thing a formal subprocessor list adds.

And the document establishing the data flow is dated 2020 and scoped to an earlier software generation, predating both the foundation model and the announced cloud collaboration, so a buyer should confirm the same separation governs the current platform.

A hosted partner algorithm also raises the accountability question this index has recorded elsewhere in pathology: when a partner model produces a finding, establish in the contract which entity carries the clinical performance claim and which one answers a complaint about it.

Performance figures come from an FDA reviewed pivotal study rather than a vendor benchmark: mean sensitivity of 97 percent, reaching 98.5 percent, and mean specificity of 98 percent, reaching 99.7 percent on some indications. The company additionally reports roughly an order of magnitude reduction in false alerts against leading single condition tools, which is the metric that determines whether a triage product survives real world use rather than being muted by radiologists. Deployment scale is reported at more than 100 million patient cases analyzed.

The published handling model is unusually specific. Before anything is uploaded, a new DICOM object is created carrying only non PHI data together with a newly generated unique identifier. Analysis runs against that object. When results return to the customer environment the identifier is used to re attach the PHI locally. Logging and analytics data are de identified before leaving as well, and PHI is encrypted at rest in the orchestrator database. Audit logs on data access are retained for three years.

The stated principle is that protected health information never leaves the customer environment, and the architecture described supports it rather than merely asserting it.

One caveat carries real weight. The document is dated 1 June 2020. A buyer should confirm the same principle governs the CARE foundation model and the current aiOS generation, where training and inference arrangements may differ from what was described for version 2.0.

The substantive answer here is architectural rather than contractual. By de identifying before upload and re attaching identifiers locally, the design limits how much protected health information crosses the boundary at all. That is a stronger control than most of what is disclosed on this axis, and it is worth more to a buyer than a policy statement would be.

What is absent is the posture itself. Aidoc does not publish its business associate agreement, does not state its role plainly in public materials, and publishes no evaluation or review cadence. The available language is that the solution is designed to comply with HIPAA and HITECH standards, which describes an intention in the design rather than an attestation or a contractual commitment.

So a buyer can see the architecture before entering a sales process but has to enter one to see the instrument.

Claims its own ISO 27001 certification with an information security management system certified annually, and publishes the control domains that system covers, including risk analysis cadence, development and production separation, patch and vulnerability management, and vendor risk management. Cloud resources are scanned for vulnerabilities quarterly and the solution is penetration tested quarterly against OWASP principles, with an additional test on major changes.

The gap is SOC 2. Aidoc does not claim its own. It cites Amazon Web Services, Google Cloud Platform and Microsoft Azure holding SOC 2 Type 2 and describes its analysis service as deployed on SOC 2 compliant infrastructure. That is an attestation covering the infrastructure provider, not an audit of Aidoc, and a buyer completing a vendor risk assessment should not record it as the latter.

No current public trust centre was found, and the certification evidence available is dated June 2020, so none of it can be verified from outside today.

The January 2026 clearance is the first FDA clearance of a double digit set of acute indications powered by a single foundation model, per the company and corroborated in trade coverage. Eleven newly cleared indications were combined with three existing into a 14 indication abdomen CT triage workflow. Positioning is precise: the company states clearly which indications are cleared, and its earlier announcement of the underlying submission described it as under review rather than cleared.

aiOS is described as providing continuous performance monitoring and built in governance for every model running on it, including models from other vendors, which is a real operational control and rare to state. Held back from A because no bias or subgroup performance evaluation and no third party AI audit were retrieved.

The grade rests on the regulatory floor rather than on anything commercial the vendor has published, and that distinction governs how every cleared vendor in this index should be read on this axis. As a manufacturer of cleared devices, Aidoc carries obligations that exist whether or not it advertises them: medical device reporting to the agency for deaths, serious injuries and malfunctions, complaint handling and corrective action under the quality system regulation, and exposure to agency compelled correction or removal.

A health system can file a report on a malfunction, and so can a patient or a clinician acting alone, through the agency's public reporting route. Reported events become searchable in the agency's public device experience database, so an institution can look up what has been reported about a product before buying it. Nothing equivalent exists for the documentation, scheduling or revenue cycle products graded elsewhere in this index, where no vendor level regulator exists at all.

The cleared indication list is also a falsifiable public commitment about what the product claims to do, reviewed against submitted performance data, and the company scopes it precisely, describing a submission as under review rather than cleared until it was. Held below the top grade because reporting is not remedy.

The route produces population level correction such as a labelling change, a field correction or a recall; it produces nothing for the individual patient whose study was affected, who has no relationship with the vendor and no standing to seek anything from it. No published indemnity toward the customer, no performance guarantee and no remediation commitment was located.

One failure mode deserves particular attention from a buyer, because it undercuts the reporting route this grade depends on. Triage reorders a worklist. A false positive is visible, since a radiologist opens the study and finds nothing. A false negative is not: the case simply is not prioritised, and the result looks identical to a normal unflagged worklist.

Harm from a missed flag therefore leaves no artifact for anyone to report, which means the volume of reports about a triage product is a weak signal of its real error rate. Ask what field monitoring detects a miss, and what the escalation path is when one is found.

Deep into the imaging workflow, thinner on the record itself. Ingests DICOM routed or queried from PACS and VNA or sent directly from modalities, plus HL7 messages from institutional sources. Results return to the desktop widget, the native worklist, the reporting system, or to PACS as a separate new series. Authentication supports Active Directory, a dedicated single sign on solution, or single sign on through PACS, RIS or dictation software, which is a practical detail that eases deployment in a radiology department.

What is not published is integration depth with Epic, Cerner or other electronic health records at the level the documentation vendors in this index disclose. For a triage product routed through imaging infrastructure that is a coherent architecture rather than a gap. A buyer who expects findings to reach the chart, or to reach clinicians outside radiology, should confirm that path separately.

Hybrid by design, and the split is documented in more detail than most vendors in this category provide. An orchestrator runs on a virtual machine inside the customer environment and handles ingestion, routing, de identification and re identification. The analysis service is a distributed cloud environment hosted on Amazon Web Services, Google Cloud Platform or Microsoft Azure. Each customer is assigned a unique storage bucket and there is no shared production storage between customers. The cloud tier provides no storage or data retention function, so uploaded data is transient rather than accumulated.

The grade is held at B for one reason. The document setting all of this out is revision 1.1 dated 1 June 2020 and is scoped to Aidoc software version 2.0 and above. It predates aiOS, the CARE foundation model and the announced multi year Amazon Web Services collaboration. A buyer should confirm that the residency split still holds for the current platform generation.

No published pricing, no published pricing mechanism and no published contract terms. Aidoc is privately held, so nothing compels disclosure the way a listing obligation does for the publicly traded vendors in this index.

One signal can read as transparency and is not. Aidoc maintains an AWS Marketplace listing, which makes the procurement channel public, but the listing states that pricing depends on the duration and terms of the contract agreed with the vendor. The route to purchase is visible; 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 vendor had published them.

Multi specialty imaging coverage spanning neuroimaging, chest, abdomen and musculoskeletal, delivered through aiOS as a single enterprise layer rather than as separate point products. The January 2026 FDA clearance consolidated 11 newly cleared indications with three existing ones into one abdomen CT workflow, 14 in total, which is unusual breadth inside a single anatomical workflow.

Read the coverage claim at the level of the cleared indication rather than the body region. Breadth here means many specific findings across several regions, not general purpose reading of any study. Products are FDA cleared and CE marked, so coverage extends to European buyers, but the indication list should be confirmed for the specific market rather than assumed from the US clearance.