An increased adenoma detection rate (ADR) reduces the risk of post-colonoscopy colorectal cancer (CRC) and death . High-quality colonoscopies are vital for preventing and treating CRC. In recent years, artificial intelligence (AI) has been introduced into colonoscopy for polyp detection (i.e. computer-aided detection [CADe]), which is expected to increase ADR according to randomized trials . However, in contrast to such promising initial results, more recent real-world studies have questioned the effect of CADe . Furthermore, a newly published article even showed a deskilling effect after the introduction of CADe for polyp detection . Nevertheless, there is still a scarcity of evidence on whether the use of CADe contributes to the improvement or deskilling of endoscopists over time. Given the unresolved questions surrounding human–AI interaction in endoscopy, the European Society of Gastrointestinal Endoscopy (ESGE) recently published a curriculum aimed at ensuring the safe and effective use of AI in this field .

To address the important question of learning effects from CADe, we evaluated how CADe use affects the polyp detection rate (PDR)-5, defined as the proportion of colonoscopies with detection of at least one polyp ≥5 mm in size; PDR-5 is a surrogate marker for ADR in Gastronet, the Norwegian national endoscopy quality registry . Gastronet utilizes PDR-5 as a convenient quality indicator for colonoscopy, as it correlates with ADR and does not require histological confirmation.

In the present study, we aimed to compare PDR-5 of endoscopists in training across three consecutive phases: 1) before CADe exposure, 2) during CADe use, and 3) after CADe removal. We also investigated the same measures for experienced endoscopists.

We included 5034 colonoscopies from 13 endoscopists who participated in all three phases ( ). After excluding 14 colonoscopies due to insufficient demographic data and 7 due to two colonoscopies being performed in the same patient, we analyzed 5013 colonoscopies performed by 7 inexperienced and 6 experienced endoscopists.

The 13 endoscopists comprised 12 physicians and 1 surgeon. At the beginning of the study, the inexperienced endoscopist group had a mean age of 41.1 years, whereas the experienced group had a mean age of 50.7 years.

Mean inclusion times from the first colonoscopy among the inexperienced vs. experienced groups were 6.8 vs. 2.2 months in phase 1, 17.7 vs. 7.6 months in phase 2, and 23.8 vs. 11.8 months in phase 3, respectively.

Patient background characteristics across the three study phases were generally similar with respect to age and sex, in both endoscopist experience groups. However, the indications for colonoscopy differed slightly, with, for example, a higher proportion of symptomatic patients in phase 3. The inexperienced group had more patients with inflammatory bowel disease (IBD) in all phases, and the experienced group had more female patients in phases 1 and 2 ( ; see also Table 1s in the online-only Supplementary material). shows the number of colonoscopies performed per endoscopist and the PDR-5 in the three phases, respectively. Most endoscopists reached at least 100 colonoscopies in phases 1 and 2, but some inexperienced endoscopists performed only a few colonoscopies in phase 3. Table 1s also shows the quality indicators in the three phases for the different experience levels.

We evaluated 5013 colonoscopies performed by seven inexperienced and six experienced endoscopists. An important finding with clinical impact was that CADe temporarily increased PDR-5 in inexperienced endoscopists, from 31.9% to 39.5% (OR 1.43, 95%CI 1.11–1.84) when CADe was used in phase 2. Another important finding was that we could not demonstrate upskilling or deskilling following CADe use. PDR-5 in non-CADe-assisted colonoscopies did not change significantly after CADe exposure, suggesting no obvious upskilling or deskilling (there was only a nonsignificant increase in PDR-5 after CAD exposure in the inexperienced group, from 31.9% to 36.3%). We observed no significant increase in PDR-5 during CAD use among experienced endoscopists. This differs from the study by Xu et al., which demonstrated an increased ADR regardless of experience . In fact, our study showed similar results for experienced endoscopists as the meta-analysis by Patel et al. and the study by Maas et al., which demonstrated no effect of CADe on ADR in a real-world setting among experienced endoscopists (nontrainees) .

The experienced endoscopists had a higher mean age than the inexperienced group (50.7 vs. 41.1 years), which may influence the effect of CADe through a conservatism bias in the human–AI interaction . We did not observe any effect of CADe among experienced endoscopists, suggesting that their skill in detecting polyps was not affected by CADe. Experienced endoscopists may underestimate, or at least do not overestimate, polyp size, whereas inexperienced endoscopists may be more prone to overestimation. The experienced endoscopists may also be less enthusiastic about a software tool intended to improve their performance, leading to an aversion to using it effectively. Algorithm aversion is addressed as a potential obstacle for human–AI interaction in the ESGE curriculum for safe and effective use of AI in endoscopy . Some endoscopists found the sound signal especially disturbing, but the box overlay did not provoke this aversion, though false-positive markings could be exhausting . Low false-positive rates are essential for CADe usability. GI Genius has to be actively turned on by the endoscopist. In the current study, two endoscopists reported having CADe active throughout the entire examination (including intubation of the colon), while the remaining endoscopists activated CADe on withdrawal from the cecum.

The finding of no significant effects of CADe on learning or deskilling among experienced endoscopists differs from the recent study by Budzyn et al., which found a significant deskilling effect among experienced endoscopists while they were exposed to CADe . A study by Troya et al. showed decreased eye travel distance when using CADe , which could explain why such deskilling happened in the study by Budzyn et al. However, there is a distinct difference in study design between our study and the Budzyn study. We evaluated the effect of exposure to CADe “after” CADe was removed, whereas Budzyn et al. investigated how endoscopists perform non-CADe-assisted colonoscopy “while” they were exposed to CADe in the other colonoscopies they performed during the same time period. Thus, it is considered that the effect of exposure to CADe will be stronger in the Budzyn study than in our study, suggesting that the deskilling effect may be a temporal phenomenon that disappears over time after CADe is removed.

Looking at the endoscopist-level analysis among the inexperienced endoscopists, three deskilled, two unchanged, and two upskilled between phases 1 and 3. Among the experienced group, three deskilled and three upskilled. This also implies that there has been neither an upskilling nor deskilling effect over time. However, it is challenging to draw such a conclusion based solely on the change in PDR-5 between phases 1 and 3, given that there is always a time effect in which endoscopists may naturally upskill or deskill over time. Mean inclusion time from the first colonoscopy among the inexperienced group was 6.8 months in phase 1, 17.7 months in phase 2, and 23.8 months in phase 3. Such long-term exposure to a medical procedure may play a role in skill development, which needs to be considered in analysis of upskilling and deskilling. This “time effect” problem applies to the recently published paper by Okumura et al. , which showed that, after CADe implementation, detection rates without CADe were maintained and did not decline over time. Among non-CADe colonoscopies, high detectors showed accelerated learning curves, indicating they maintained a higher ADR, whereas low detectors had no significant change in their learning curves. However, the inclusion period was quite long (3 years), compromising the study interpretation due to the possible “time effect” on learning.

In all phases, the inexperienced endoscopists had a higher PDR-5 than the experienced endoscopists. This may be related to inexperienced physicians overestimating polyp size or to their more meticulous mucosal evaluation than that of experienced physicians. The median withdrawal time across all phases was 10 minutes (IQR 8–16) for the inexperienced endoscopists and 9 minutes (IQR 7–11) for the experienced endoscopists. The withdrawal time must be evaluated with caution, however, as many examinations were excluded because of biopsies/therapy. Both experience groups had longer withdrawal times during the CADe phase. Other studies have shown similar results, with a slight increase in the withdrawal time using CADe . A study by Schult et al. showed that trainees detected more adenomas than consultants . Following an initial training period of 300 colonoscopies, ADR among trainees remained unchanged . This finding aligns with the study by Gianotti et al. , which demonstrated an increase in ADR after trainees performed more than 140 colonoscopies under supervision, and subsequently outperformed consultants. Thresholds of 325 colonoscopies in male patients and 539 colonoscopies in female patients were determined to be ideal for achieving adequate ADR. In the current study, we defined an experienced endoscopist as having performed ≥1000 colonoscopies during their career at the start of study inclusion. This may be an adequate cutoff when comparing our findings to those in these two previous studies . Our findings that inexperienced endoscopists outperform experienced endoscopists also align with these studies.

A strength of our study is the large number of colonoscopies performed by 13 endoscopists, both inexperienced and experienced. The more procedures included per endoscopist, the greater the reliability of the evaluation of changes in PDR-5. We increased the number of included endoscopies to approximately 5000 after the meta-analysis by Patel et al. showed no significant effect of CADe among experienced (nontrainee) endoscopists in a real-world setting . Another strength is the three-phase design, in which all endoscopists included colonoscopies in all phases, serving as their own controls. The colonoscopies were included at three different levels of outpatient care: university hospital, local hospital, and high-volume outpatient endoscopy practice. The design reflects a variety of endoscopy settings and experience levels, making this study broadly relevant for real-world endoscopy.

There are several limitations to this study. The most important is the use of PDR-5 as a quality indicator for lesion detection, rather than the gold standard ADR. A meta-analysis by Niv et al. found that a ratio of 0.68 can be used to estimate ADR from PDR for the individual endoscopist or a group of endoscopists before receiving formal results from the pathology department . PDR in colonoscopies is higher than PDR-5, suggesting that PDR-5 may have a ratio closer to 1 compared with PDR. PDR-5, being measured optically, may lead to over- or underestimation of polyp size, and studies have shown significant interphysician variation for estimating polyp size . To collect the large number of colonoscopies (approximately 5000), we decided to use PDR-5. We acknowledge that this quality indicator is less well-validated than ADR. Nevertheless, PDR-5 has been used for many years in Norway, and previous studies have adopted it . Another consideration was our aim to conduct a pragmatic study, in which endoscopists performed examinations according to their usual practice. Although polyps are sent to histology, we use PDR-5 as our quality indicator because it is both readily available and independent of histology. Notably, a strength of using PDR-5 is that it includes both adenomas and sessile serrated lesions, and the latter may account for a substantial proportion of interval cancers . It should be noted that this study may not fully capture the effect of CADe on polyp detection when only polyps ≥5 mm are recorded. However, some criticism of CADe is that it only increases the detection of diminutive polyps with very low malignancy potential. Focusing on polyps ≥5 mm may thus reveal whether CADe affects the more relevant polyps. Future studies are needed to validate PDR-5 as a colonoscopy quality indicator.

Some inexperienced endoscopists performed relatively few endoscopies in phase 3, which may have led to underpowering of phase 3. This may have limited the ability to detect modest sustained effects of CADe. Future studies are needed to evaluate the learning effect of CADe among inexperienced endoscopists. As shown in previous studies and in the same region of Norway , patient age, sex, and colonoscopy indication significantly influence ADR/PDR-5. To adjust for this effect, we performed multivariable analyses. A possible limitation of the multivariable analyses is the risk of underpowering. We might have overlooked otherwise significant differences due to limited sample size. In particular, the inexperienced endoscopists had differences in the case-mix between phases 1 and 3, with higher age, more women, more symptoms, and fewer IBD patients in phase 3. This difference in case-mix may have influenced the PDR-5. Many of our patients in routine practice have the indications of symptoms and IBD, respectively. Other studies exclude patients with IBD, but we decided to include this indication as it reflects real-world data. However, it is a limitation that patients with IBD may exhibit different mucosal dysplasia features, which may not be detected by the training set used in GI Genius . Another limitation to the generalizability of this study is that the definition of an experienced endoscopist is not well established in the literature. Finally, we used only one type of AI device, which may limit the generalizability of our results. However, a study comparing CADe devices from different companies showed only minor differences between the producers .

Overall, these results support a targeted deployment strategy for CADe, prioritizing its use among trainees and early-career endoscopists, where a clear PDR-5 benefit is observed. The added value for high-volume, experienced endoscopists in real-world practice may be limited. Additionally, the absence of an observable post-exposure reduction in PDR-5 compared with pre-exposure argues against the idea that exposure to CADe erodes skills, at least for nonexperienced endoscopists, although such skill reduction has been observed in another study involving only experienced endoscopists .