18 November 2026 · 13 min

Company Spotlight: DigestAID Deep Capsule - Sifting Millions of Frames for GI Bleeds

Radiology & ImagingFDA Clearances

DigestAID recently secured FDA 510(k) clearance for Deep Capsule, an AI tool designed to help gastroenterologists review capsule endoscopy footage for small bowel bleeding. Maya and Sam unpack the clearance summary to explore the algorithm's performance data, its substantial equivalence to a predicate device, and the real-world trade-offs between high sensitivity and low specificity when reviewing millions of video frames.

Key points

Source: FDA 510(k) summary K250655: Deep Capsule® (Deep Capsule US) (Digestaid - Artificial Intelligence Development) - U.S. Food and Drug Administration, 2026

This episode is an AI-generated conversation summarising a public document; the hosts' voices are synthetic. It is for information only and is not medical advice. Always refer to the original source.

Company: Digestaid - Artificial Intelligence Development (digestaid.health) · Product: Deep Capsule® (Deep Capsule US) · FDA 510(k) K250655

Your company here. This podcast is looking for its first sponsors: reach clinicians, health-system leaders and medtech and pharma teams following AI in medicine. Sponsorship options and rates →

Transcript

Maya: Welcome to AI in Medicine - Smart Summaries. I am Maya, and today we are looking at an AI tool that took the diagnostic yield for finding small bowel lesions from 0.761 with standard clinical reading to a massive 0.961 with AI assistance.

Sam: I am Sam. Just a quick reminder before we dive in, our voices are AI-generated, and this episode is a summary of a public document. Today we are doing a Company Spotlight on a product called Deep Capsule, based entirely on the FDA 510(k) summary K250655 published by the U.S. Food and Drug Administration.

Maya: This document outlines the clearance for DigestAID, a company based in Portugal, to market Deep Capsule in the United States. It is a Class II gastrointestinal capsule endoscopy analysis software device designed to help clinicians review video footage for adult patients with suspected small bowel bleeding.

Sam: Let us pause right there and set the stage for anyone who does not sit in a gastroenterology clinic all day. Capsule endoscopy is exactly what it sounds like. A patient swallows a pill with a tiny camera inside, and it takes pictures as it travels through the digestive tract.

Maya: Exactly. And it takes a lot of pictures. We are talking about hours of video footage per patient, generating tens of thousands of frames. A human reviewer then has to sit and stare at this footage, looking for a tiny source of bleeding in the small bowel. It is tedious, it causes immense reader fatigue, and it creates massive backlogs in clinical workflows.

Sam: Which is exactly why companies are building AI to help. So, what exactly does Deep Capsule do with all that video? According to the summary, it is a computer-aided detection tool, or CADe, that uses deep learning and convolutional neural networks to find potential lesions.

Maya: That is right. It does not differentiate the lesions, meaning it will not tell you if it is an ulcer or a tumor. It just flags that a lesion might be there. The software pulls these suspected findings automatically from the video frames and displays them on the right side of the screen as a structured gallery of thumbnails.

Sam: So it basically creates a highlight reel of the most suspicious frames for the doctor to look at.

Maya: Yes, but the FDA summary is incredibly clear on the limitations. This is not for autonomous diagnostic use. The document states explicitly that a negative result determined by the device alone does not exclude small bowel disease, which would be a false negative. And a positive result does not automatically confirm disease, a false positive. The ultimate diagnosis relies on a human in the loop.

Sam: Right, the clinician still has to review the complete video in accordance with standard clinical practice, especially if the patient's symptoms persist. Now, because this is a 510(k) clearance, DigestAID had to prove that Deep Capsule is substantially equivalent to a predicate device already on the market.

Maya: They did. They compared it to the NaviCam ProScan, which was cleared under a De Novo pathway. Both use AI to detect potential small bowel lesions to decrease review time. But there is one functional difference noted in the comparison table.

Sam: I saw that. The NaviCam ProScan also assists reviewers in identifying the anatomical location of the digestive tract, like telling you if the image is from the oral cavity, esophagus, stomach, or small bowel. Deep Capsule does not have that location feature.

Maya: Correct. Deep Capsule focuses purely on the small bowel and identifying potential lesions. To prove this works, DigestAID ran the algorithm through several phases of testing, starting with standalone bench performance.

Sam: Let us dig into those numbers. The initial dataset for training and standalone testing was built from 1,133 patients who had capsule endoscopies at 2 clinical institutions in Portugal. How many images are we talking about here?

Maya: A staggering 321,357 expert-labeled images. They split the 1,133 patients into two groups. They used 861 patients, or 76.0% of the dataset, to train the model. Then they took the remaining 272 patients, making up 24.1%, and used them to test the model internally.

Sam: And how did it do on that internal test of 272 patients?

Maya: At the patient level, the sensitivity was 95.8% and the specificity was 75%. When you break it down to the individual image level, analyzing normal versus abnormal frames, the sensitivity was 92.1% and the specificity was 88.0%.

Sam: Okay, 95.8% sensitivity at the patient level is very high. It means it is very good at catching patients who actually have lesions. But an internal test on data from just 2 clinical institutions is one thing. The FDA wants to see how it performs across different clinical settings. That brings us to the clinical validation study.

Maya: Yes, the clinical validation was a retrospective study involving 330 capsule endoscopy exams collected between January 2021 and April 2024. These came from seven independent clinical centers across four countries. The United States, Portugal, Brazil, and Spain.

Sam: And none of these sites overlapped with the training data, right?

Maya: Exactly, no overlap. The study involved 15 board-certified gastroenterologists who acted as the clinical readers. They were compared against a reference standard set by an independent expert board of 5 adjudicators.

Sam: So how did they run the test? Did the doctors read it with and without the AI?

Maya: It was done in two phases. First, experts at the originating centers interpreted the videos following standard of care reading procedures. Then, in the second phase, independent readers who were blinded to the initial reports and patient data used the Deep Capsule AI software to assist their review.

Sam: This is where we get back to those numbers you mentioned at the very beginning. The diagnostic yield. This is the proportion of patients where at least one clinically relevant lesion was detected.

Maya: Right. Out of the 330 total patients, the expert board ground truth established that 290 patients actually had lesions. The standard of care reading only found lesions in 251 patients, which is a diagnostic yield of 0.761. But the AI-assisted reading found lesions in 317 patients, pushing the diagnostic yield to 0.961.

Sam: Wait, really? The AI arm found 317 positive cases, but the expert board ground truth was only 290? That means the AI plus human team was calling more positive cases than actually existed.

Maya: Yes, and that is a crucial point for anyone looking to buy or use this tool. Let us look at the raw estimates for sensitivity and specificity in this clinical validation. The AI-aided sensitivity was 0.972, compared to 0.762 for the standard of care. So it is catching way more real lesions.

Sam: But there is always a trade-off. If sensitivity goes up that much, what happened to the specificity?

Maya: The specificity for the standard of care was 0.250. For the AI-aided arm, it dropped to 0.125.

Sam: Wow. 0.125 is extremely low. That implies a very high false-positive rate. But the document mentions something about the dataset prevalence affecting this, right?

Maya: Yes, it does. The summary notes that specificity estimates varied across subgroups because of the high prevalence of positive cases in the clinical dataset. Remember, out of 330 exams, 290 were positive. That left only 40 true negative patients in the entire study to test specificity against.

Sam: That makes sense. If almost everyone in the test has a lesion, it is harder to get a precise measure of how well the tool ignores a healthy patient. Still, practically speaking, if I am a hospital administrator buying this, I have to know that this tool errs on the side of caution. It is going to flag a lot of things as lesions that are actually perfectly normal.

Maya: Exactly. The model prioritizes sensitivity. It is designed so that you do not miss a bleed. The trade-off is that the clinician will have to spend time reviewing those false positives in the thumbnail gallery. However, the summary concludes that despite this, the mean reading time was significantly reduced with AI plus physician reading compared to standard reading.

Sam: It is interesting that the FDA summary uses the words significantly reduced, but it does not actually give us the exact time reduction. I would love to know exactly how many minutes are saved per exam.

Maya: The document does not provide the exact time reduction, so we have to leave it at that. But it does provide an incredible look at the scale of the image-level analysis in that clinical validation. They evaluated 4,175,392 video frames.

Sam: All 4,175,392 frames manually reviewed by the experts to establish ground truth. That is a massive undertaking.

Maya: And on those 4,175,392 frames, the AI-aided reading identified 813,203 as lesion frames, and 3,362,189 as normal frames. At the image level, the sensitivity was 94.0% and the specificity was 84.9%.

Sam: Okay, that frame-level specificity of 84.9% sounds much better than the patient-level specificity of 0.125. It means that frame by frame, it is actually pretty good at knowing what is normal.

Maya: Yes, but remember, it only takes a few false-positive frames to flag a patient as positive for a lesion, which explains the difference between the frame-level and patient-level metrics.

Sam: Let us talk about hardware compatibility. Capsule endoscopy systems are proprietary. Did they test this across the board, or just on one brand?

Maya: This is a vital point for procurement teams. The clinical validation was heavily weighted toward one device. Out of the 330 patients, 307 used the PillCam SB3 system. They also had 21 patients using the Olympus EC-10, and just 2 using the PillCam SB2.

Sam: Just 2 patients on the SB2? That is barely a blip. What about other brands on the market?

Maya: The document explicitly states that performance metrics for certain systems like the PillCam SB2 are based on limited sample sizes and should be interpreted accordingly. They also mention a device called the OMOM HD, which was included in internal development but is not part of the U.S. compatibility claims.

Sam: So if your health system uses the PillCam SB3, you have a solid block of validation data here. If you use something else, you need to look very closely at the labeling to ensure Deep Capsule is compatible and validated for your specific hardware.

Maya: Exactly. They also looked at demographic subgroups. The clinical validation dataset included 179 females and 151 males. They broke the data down by age ranges, and they also looked at race and ethnicity, including White, Black, Hispanic, Asian, and others.

Sam: Did the AI struggle with any particular demographic or clinical indication?

Maya: According to the summary, no clinically meaningful degradation in overall performance was observed across major demographic or technical subgroups. The AI consistently demonstrated higher sensitivity compared to the standard of care across the board.

Sam: That is reassuring. The indications for the procedure varied too, right? People get capsule endoscopies for different reasons.

Maya: Yes, they listed several indications in the 330 patients. Half of them, 166 patients, presented with anemia. Then there was Crohn's disease at 72 patients, and obscure gastrointestinal bleeding at 67 patients. They even had a handful of patients with diarrhea, genetic syndromes, and celiac disease.

Sam: And the performance held up across those different clinical presentations.

Maya: It did. Ultimately, the FDA found that Deep Capsule is substantially equivalent to the NaviCam ProScan and does not raise different questions of safety and effectiveness.

Sam: To wrap this up, if I am a clinician or a health system leader looking at this, my one memorable takeaway is this: Deep Capsule is a highly sensitive net for catching potential small bowel lesions. It took sensitivity from 0.762 up to 0.972. It significantly reduces overall mean reading time, but the FDA label requires you to still review the complete video, and you must be prepared to manually sift out the false positives from that thumbnail gallery.

Maya: Perfectly said. It is an assistant, not a replacement for clinical judgment. And if you are buying it, check that your capsule hardware matches what was validated, primarily the PillCam SB3.

Sam: That is it for today's episode. A crisp reminder that you can find a link to the full FDA 510(k) summary in our show notes. And as always, this podcast is for informational purposes only and is not medical advice.

Maya: Thanks for joining us on AI in Medicine - Smart Summaries. See you next time.