Deep Capsule (Deep Capsule US)

Performance Metrics

Clinical Evidence

Two bodies of evidence need to be kept apart. The first is the device evidence the FDA reviewed for K250655, summarised in the public 510(k) summary. The second is a run of peer-reviewed papers from the Porto gastroenterology and engineering group that founded DigestAID; none of those papers names Deep Capsule, and the models they describe differentiate lesion types and grade bleeding risk, which the cleared US software does not. They are background research, not performance data for the product on sale. The FDA-reviewed standalone test used 272 patients held out from a 1,133-patient, 321,357-image dataset collected at two Portuguese centres. Patient-level sensitivity was 95.8 percent (95 percent CI 92.7 to 97.9) and specificity 75 percent (95 percent CI 34.9 to 96.8), with an AUC of 0.966. The specificity figure rests on eight lesion-free patients, hence the very wide interval; the FDA summary itself notes that clinicians will need to review the findings carefully to identify false positives. Image-level sensitivity and specificity on 101,802 frames were 92.1 and 88.0 percent. The clinical validation is a retrospective multicentre reader study of 330 capsule examinations recorded between January 2021 and April 2024 at seven centres in Portugal, Spain, Brazil and the United States (118 US exams from a single site), read by 15 gastroenterologists, with a separate expert board as reference standard. Diagnostic yield, defined as the proportion of patients with at least one clinically relevant small-bowel lesion, was 96.1 percent (93.4 to 97.8) for AI-plus-physician reading against 76.1 percent (71.2 to 80.3) for standard reading and 87.9 percent for the expert board; non-inferiority was established (p below 0.001). Per-patient sensitivity rose from 76.2 to 97.2 percent, while specificity fell from 25 to 12.5 percent on 40 lesion-free patients, so the gain is bought with more positives for the reader to dismiss. The summary reports mean reading time as significantly reduced without stating the figures; the published version of the study gives a mean AI-assisted reading time of 203 seconds per examination. That published version, Mascarenhas Saraiva et al. in the American Journal of Gastroenterology (online August 2025, print April 2026), describes the same 330-exam, seven-centre, four-country cohort and reports a lesion detection rate of 96.1 versus 76.3 percent and sensitivity of 97.5 versus 78.2 percent, with minor differences from the FDA-reviewed figures. Two cautions: the abstract does not name the product and reports a model that also differentiates lesions, and it describes the design as prospective where the FDA summary calls the study retrospective. This entry follows the FDA description. The three earlier Porto papers (BMJ Open Gastroenterology 2021; Endoscopy International Open 2024; GE Portuguese Journal of Gastroenterology 2024) report frame-level convolutional-network results for lesion differentiation, Saurin bleeding-risk stratification and multi-brand vascular-lesion detection. They are listed under publications as research models. Joao Ferreira, the engineering lead, is disclosed as a paid DigestAID employee in both 2024 papers; the 2021 paper declares no competing interests and predates the company's founding. The honest gap is not the absence of a reader study but its provenance and design: every study is manufacturer-run, the validation is retrospective, independent validation of the cleared software has not been published, and no evidence addresses use outside the suspected-small-bowel-bleeding indication.

Inside the Auris+ Listing

Regulatory Approvals

Safety Record

No FDA recall or adverse-event cluster attributable to Deep Capsule was located in public reporting as of the review date; the FDA enforcement and MAUDE databases could not be queried directly from this environment, and the device's US market history only began in March 2026, so the post-market signal is short by construction. The harm surface of the category is indirect: the software selects frames for the clinician rather than issuing a diagnosis, so the principal risk is a missed lesion in frames the reader de-prioritises, and the labelling carries an explicit warning against over-reliance. The FDA-reviewed data describe a sensitivity-first profile: patient-level standalone sensitivity of 95.8 percent, and in the reader study 8 of 290 expert-board-positive patients missed by AI-assisted reading against 69 missed by standard reading. The cost is on the other side of the ledger: standalone specificity of 75 percent on eight lesion-free patients and AI-assisted per-patient specificity of 12.5 percent mean a large share of flagged frames will be false positives that the reader must clear. Procurement reviewers should also note the category convention: the clinician remains responsible for the full study, and the software's output is an attention aid, not a screen-out.

Intended Use & Indications

Deep Capsule is a software-only reading aid for small-bowel capsule endoscopy. Its FDA-cleared indication (K250655, decided 12 March 2026) reads, in the 510(k) summary's wording, as an artificial-intelligence-assisted reading tool designed to aid small-bowel capsule-endoscopy reviewers in decreasing the time to review capsule-endoscopy images for adult patients in whom the images were obtained for suspected small-bowel bleeding, with the clinician responsible for conducting their own assessment through review of the entire video. The scope is deliberately limited. The same summary states that the device is capable of detecting small-bowel lesions without differentiating them, that it is not intended to characterise lesions in a manner that would replace biopsy or other characterisation tools, and that it is not a stand-alone diagnostic. Its AI output is a gallery of suspected-lesion frames selected from the video; a findings count and exam metadata are non-AI outputs. The device belongs to the FDA's young capsule-endoscopy analysis software category (product code QZF, 21 CFR 876.1540, established by the NaviCam ProScan De Novo), in which the clinician remains the reader of record. Compatibility is limited to the small-bowel capsule systems named in the labelling.