Company Spotlight: Aidoc BriefCase-Triage - Triage Time Cut to 1.09 Minutes

Maya: Imagine a software tool cutting the time it takes to notify a clinician about a suspected vessel occlusion down to exactly 1.09 minutes.

Sam: That is the actual data point we are looking at today. Welcome to AI in Medicine, Smart Summaries. Today we are diving into a Company Spotlight episode focused entirely on the FDA 510(k) summary for Aidoc's BriefCase-Triage, which was published by the U.S. Food and Drug Administration on May 14, 2026.

Maya: Before we begin, we want to remind everyone that our voices are AI-generated, and we are unpacking a publicly available FDA 510(k) summary, not company marketing material. I am Maya, looking at this from the clinical and health-system perspective.

Sam: And I am Sam. This is a very specific type of document, Maya. A 510(k) clearance is based on substantial equivalence to a predicate device, as the summary itself states. Our goal here is to be completely balanced and independent. So, what exactly is BriefCase-Triage?

Maya: BriefCase-Triage is categorized under product code QAS, which stands for Radiological Computer Aided Triage And Notification Software. It is a Class II device. The software is based on a deep learning AI algorithm programmed component and is intended to run on a linux-based server in a cloud environment.

Sam: And who is this for? What is the actual indication for use?

Maya: It is indicated for use in the analysis of head CTA images in adults or transitional adolescents aged 18 and older. The device is intended to assist hospital networks and appropriately trained medical specialists in workflow triage by flagging and communicating suspected positive findings.

Sam: Does the document specify which specific vessels it looks at? Because that matters quite a bit for clinical workflow and what the care teams can actually expect it to find.

Maya: It does, very specifically. For complete Large Vessel Occlusions, or LVOs, it covers the MCA-M1, PCA-P1, ACA-A1, ICA, and the Basilar artery. For Medium Vessel Occlusions, or MeVOs, it covers the MCA-M2, MCA-proximal M3, PCA-P2, PCA-proximal P3, ACA-A2, ACA-proximal A3, and Vertebral-V4.

Sam: That is a comprehensive list of vessels. But how does it physically work for the radiologist sitting in the reading room? How does this information actually reach them?

Maya: The summary breaks down the workflow step by step. The software receives filtered DICOM images and processes them chronologically. It runs the algorithms on each series to detect suspected cases. Following the AI processing, the output of the algorithm analysis is transferred to an image review software, which is a desktop application.

Sam: So it sits on a cloud server, does the math, and then pushes a notification to a desktop app. What does the user see when a case is flagged?

Maya: When a suspected case is detected, the user receives a pop-up notification. They are presented with a preview function. But the summary is very strict about what this preview is. It is a compressed, low-quality, grayscale image. It is explicitly captioned with the words 'not for diagnostic use, for prioritization only'.

Sam: Wait, really? It gives them a low-quality image on purpose?

Maya: Yes, it is basically a safety and compliance feature. The document states this preview is meant for informational purposes only. It does not contain any marking of the findings, and it is not intended for primary diagnosis beyond notification. The device does not alter the original medical image.

Sam: So the low quality forces the clinician to go back to the standard, high-resolution PACS system to do their actual job and make the diagnosis. It is just a tap on the shoulder.

Maya: Exactly. The document states that presenting the users with worklist prioritization facilitates efficient triage by prompting the user to assess the relevant original images in the PACS. It also notes that the device operates in parallel with the standard of care interpretation.

Sam: Does it rearrange the existing worklist? If I am reading a stack of non-urgent scans, does it physically bump the suspected stroke to the top of my queue?

Maya: It does not remove cases from the standard of care queue, and it does not de-prioritize cases. The summary specifically says it does not disturb standard interpretation of the images. It operates in parallel. It just gives you that separate pop-up notification so the suspect case receives attention earlier than it would have otherwise.

Sam: You mentioned earlier that this is a 510(k) clearance, which requires demonstrating substantial equivalence to a predicate device. What is the predicate device they compared this to?

Maya: The primary predicate device is Aidoc's own Briefcase for VO, which was cleared under document number K220709. The summary actually notes that the subject device and the algorithm analysis module for the predicate are identical.

Sam: If the algorithm is identical, why did they need to go through the process of getting a whole new FDA clearance?

Maya: The difference is the inclusion criteria for the clinical performance testing. The predicate device, K220709, was cleared for head CTA images. This new clearance, K261317, adds mCTA images to the inclusion criteria.

Sam: So they expanded the types of scans the algorithm is officially allowed to process. Did they test it specifically on those new mCTA scans to prove it actually works?

Maya: Yes, they did. Aidoc conducted a retrospective, blinded, multicenter study designed to evaluate the software's accuracy in flagging suspected mCTA vessel occlusion findings. The primary objective was to evaluate the accuracy of the software in analysis and identification of vessel occlusion findings in head CT images using the mCTA acquisition protocol.

Sam: Whenever we talk about retrospective studies, we have to point out that it means they used historical data, not a live, real-world clinical trial going forward. It is a limitation listeners should always keep in mind when evaluating AI. How did they determine what was truly a positive or negative case in that historical data? Who decided the ground truth?

Maya: The accuracy was determined by comparing the software's positive or negative determination against a ground truth. That ground truth was established by what the document calls 2+1 expert, U.S. board-certified radiologist reviewers using majority voting.

Sam: So two radiologists review it independently, and if they disagree, a third breaks the tie. That is a solid, standard methodology for this kind of study. Let us get to the real numbers. How did the algorithm perform on its primary endpoint?

Maya: The primary endpoint was met. The sensitivity was 81.4% and the specificity was 85.5%.

Sam: Let us break that down for the boardroom. Sensitivity at 81.4% means it is catching a high rate of actual vessel occlusions in this dataset. Specificity at 85.5% means it is correctly identifying the true negatives effectively too. The document notes that both of these exceed the 80% sensitivity and specificity thresholds.

Maya: That is correct. But the summary also includes secondary endpoints, which cover the predictive values. The Negative Predictive Value, or NPV, was 97.6%.

Sam: That is very high. A 97.6% NPV means if the software says there is no vessel occlusion, it is almost certainly right based on this study. That should give clinicians a lot of comfort. What was the Positive Predictive Value?

Maya: The Positive Predictive Value, or PPV, was 38.4%.

Sam: Wait, 38.4%? That means if I am a radiologist and this system flags a scan with a pop-up notification, it is a true positive only 38.4% of the time according to this testing. That is a significant number of false positives.

Maya: It is a significant number of false positives. But this brings us back to the intended use. It is a triage tool. It does not diagnose. It highlights cases so the radiologist looks at them sooner in the PACS. The trade-off for catching time-sensitive critical cases earlier is accepting that some flags will be negative upon formal review.

Sam: That is a crucial point for implementation. If you are a hospital buyer rolling this out, you have to prepare your clinical team for that reality. They will get alerts for cases that turn out to be negative. Speaking of speed and catching things earlier, what were the time-to-notification results?

Maya: This is the hook we mentioned at the very start. The secondary endpoint included a time-to-notification analysis. The summary defines this as the flagged scan details being input into the Aidoc worklist. That time was calculated as 1.09 minutes.

Sam: And how does that 1.09 minutes compare to the predicate device? Because speed is the entire value proposition here.

Maya: The summary explicitly compares it to the predicate device, noting that the predicate had a time of 2.23 minutes.

Sam: Bringing the processing time down from 2.23 minutes to 1.09 minutes is a notable reduction when you are dealing with a potential stroke or vessel occlusion. Every single minute literally translates to brain tissue. What about the likelihood ratios?

Maya: The summary lists the Positive Likelihood Ratio as 5.6. The Negative Likelihood Ratio was 0.22.

Sam: Let us talk about the data they excluded from the testing. Were there specific types of scans the AI was not evaluated on?

Maya: Yes, there is a very interesting comparison in the summary table regarding inclusion and exclusion criteria. For the new subject device, the inclusion criteria required head CTA or mCTA images performed on adults or transitional adults 18 years of age and older, with a slice thickness between 0.5 mm and 1.25 mm. The only exclusion listed is all studies that have an inadequate field of view.

Sam: And how does that compare to the predicate device's criteria?

Maya: The predicate device had stricter criteria. It required a head CTA protocol with a 64-slice scanner or higher. And for exclusions, the predicate excluded all scans that are technically inadequate, including motion artifacts, severe metal artifacts, suboptimal bolus timing, or an inadequate field of view.

Sam: That is a major difference. The new subject device does not list motion artifacts, severe metal artifacts, or suboptimal bolus timing as exclusions for the clinical performance testing. Does the document explain why they dropped those specific exclusions for this test?

Maya: It does not provide any explanation for that change. It simply lists the final inclusion and exclusion criteria side by side in Table 1.

Sam: That is exactly the kind of thing a health-system IT buyer or a radiology chief might want to ask the company about directly. Speaking of IT, we know this is a software component running on a linux server in a cloud environment. Hospital IT departments are incredibly cautious about opening up their imaging networks to the cloud. What does the summary say about security?

Maya: There is a dedicated section specifically for Cybersecurity. It states that cybersecurity has been incorporated into the software development lifecycle in alignment with Section 524B of the FD&C Act and FDA cybersecurity guidance.

Sam: What specific measures did they document taking?

Maya: The summary states Aidoc implemented a risk-based approach to cybersecurity. This includes secure design practices, vulnerability assessments, a Software Bill of Materials, often called an SBOM, and penetration testing. The document notes these efforts support the safety, effectiveness, and resilience of the software against cybersecurity threats.

Sam: An SBOM is basically an ingredient list for the software. So if a vulnerability is found in a piece of open-source code months later, the hospital knows if they are exposed. That is becoming a standard requirement, but it is excellent to see it formally documented here. What else does the clearance letter itself tell us about post-market requirements?

Maya: The FDA letter to Aidoc mentions several ongoing requirements. It cites the Quality Management System Regulation, 21 CFR Part 820. It reminds the company they are subject to ISO 13485 for design controls, nonconforming product, corrective action, and preventative action. It notes they must document changes and approvals in the Medical Device File.

Sam: The letter also mentions the Unique Device Identification System rule, or UDI Rule. That means the device has to bear a unique device identifier on its label, and certain information must be submitted to the Global Unique Device Identification Database.

Maya: Correct. And one more detail on how the AI was built: The summary notes that as is customary in the field of machine learning, deep learning algorithm development consisted of training on labeled, or tagged, images. Each image in the training dataset was tagged based on the presence of the critical finding.

Sam: So, stepping back and looking at the big picture for this clearance. What is the memorable takeaway here for clinicians and health systems?

Maya: The big takeaway is speed and workflow expansion. They proved substantial equivalence to their predicate device while expanding the software to analyze mCTA images. They documented an impressive time-to-notification of 1.09 minutes, down from 2.23 minutes. The trade-off is the 38.4% Positive Predictive Value, meaning teams need to be prepared for false alarms, but the high Negative Predictive Value of 97.6% shows it is highly reliable when it clears a scan.

Sam: And on the workflow side, it is crucial to remember this does not disrupt the standard queue. It pops up a low-quality, compressed, grayscale preview image purely to say, 'Hey, look at this one next on your main workstation.' It is a parallel triage assistant, not an automated diagnostician.

Maya: Well said. That wraps up our deep dive into the FDA 510(k) summary for Aidoc's BriefCase-Triage, clearance number K261317.

Sam: As a reminder, you can find a link to the public source document in our show notes. We highly encourage you to read the full summary for yourself. This podcast is for informational purposes only and is not medical advice. Thanks for listening.