Capsule endoscopy (CE) is a non-invasive diagnostic modality that has become a valuable tool for evaluating small bowel (SB) pathology. Despite its clinical value, it is limited by the time required for complete video interpretation and the associated reader fatigue. Studies have shown that accuracy drops even in the hand of experts after just one video. Artificial intelligence software has been developed to assist the reporting process. TOP100 is an integrated software that selects 100 images most likely to contain abnormalities (1). The aim of this study was to assess the overall diagnostic performance of TOP100 for specific SB findings in real-life clinical practice, by comparison with the standard reader together with its performance across different clinical indications. 

A retrospective single-centre cohort study was conducted at our tertiary referral centre. We included all consecutive patients who underwent CE using the PillCam™ SB3 system between January 2024 to August 2025.  Data collected included: baseline demographics, comorbidities, previous endoscopy and laboratory findings, indications for CE, adequacy of bowel preparation.  CE findings were compared between the standard reader (SR) and a second reader using the TOP100 images, who was blinded to the original report. All CE readers had read >500 CE in their lifetime.

The sensitivity, specificity, positive predictive value (PPV) and negative predictive value (NVP) of TOP100 were compared to the SR (gold standard) for each type of CE finding (inflammatory, bleeding etc). A sub-analysis of diagnostic accuracy was also performed based on the CE indication.  

A total of 812 patients were included, however 5 were excluded from the analysis, with 3 unable to swallow the capsule and 2 lacking available Top100 readings. The final analysis comprised of 807 patients (57% female, median age 53 years, IQR 34–65). The most common indications were suspected or known inflammatory bowel disease (IBD) (50%), and iron-deficiency anaemia (24%). In 88% of patients the bowel prep was adequate and 90% had a complete examination. Median SB transit time was 3h51mins (2h46 – 4h50m).

The diagnostic yield of findings at CE was significantly higher by SR compared to TOP100 (40% vs 29%, p<0.01). For any indication, the sensitivity and specificity of TOP100 for active bleeding was 58% (54%-61%) and 97% (95%-98%), and for angiodysplasias was 58% (54%-61%) and 97% (95%-98%), respectively. In patients with overt bleeding, TOP100 identified 80% of the total angiodysplasias and 71% of active bleeding reported by SR, with a sensitivity and specificity for P2 lesions of 65% (59%-77%) and 100% (100%-100%), respectively. The overall sensitivity and specificity of TOP100 for ulcers was 53% (49%-56%) and 96% (94%-97%), and for erosions was 40% (40%-44%) and 90% (88%-92%). In patients with suspected or known IBD, the diagnostic performance of TOP100 for ulcers was similar, while sensitivity improved for erosions by almost 12%.  

SR remains the reference standard for reading and reporting CE, however TOP100 can be a useful tool for quick preview of the video to assist in capsule reading and identifying cases that needs to be prioritized – especially in high volume centres. The diagnostic performance is also dependent on the indication of CE with improved performance noted in bleeding and IBD.