Download PDF
Endoscopy
DOI: 10.1055/a-2710-3248
Editorial

Long-term learning and not deskilling: a reassuring signal for computer-aided detection-assisted colonoscopy

Referring to Okumura T et al. doi: 10.1055/a-2661-2624

Authors

  • Pieter Sinonquel 

    1   Department of Gastroenterology and Hepatology, University Hospitals Leuven, Leuven, Belgium (Ringgold ID: RIN60182)
    2   Translational Research in Gastrointestinal Diseases (TARGID), KU Leuven Biomedical Sciences Group, Leuven, Belgium (Ringgold ID: RIN71510)
Further Information
Also available at
   Permissions and Reprints

10.1055/a-2661-2624

In the rapidly evolving field of gastrointestinal endoscopy, the integration of artificial intelligence (AI), particularly computer-aided detection (CADe), has shown promise in enhancing the adenoma detection rate (ADR), one of the key quality indicators of screening colonoscopy [1] [2] [3]. Yet, as enthusiasm for AI grows, so do the concerns about its unintended consequences. Foremost among these is the possibility of “deskilling,” a phenomenon in which over-reliance on AI, referred to as automation bias, might erode an endoscopist’s diagnostic and even technical performance over time.

To address this unexplored backdrop, the prospective observational study by Okumura et al. [4] offers a reassuring and well-described perspective on what the impact of CADe may be on endoscopists’ learning curves. The authors investigated, for the first time, the long-term evolution of the ADR among endoscopists exposed to CADe. With over 4700 propensity score-matched colonoscopy cases and follow-up for 3 years, the authors evaluated performance trends using cumulative summation analysis, a robust and objective tool to monitor real-time skill acquisition and retention, which is independent of sample size.

“Okumura et al. provide the first robust evidence that long-term CADe use does not deskill endoscopists and may in fact contribute to meaningful learning gains, albeit particularly among already proficient detectors.”

Their conclusions are both lucid and important: their study did not reveal any evidence of deterioration in ADR in procedures performed without CADe after its introduction into clinical practice. This directly challenges the hypothesis that CADe leads to deskilling over time. In contrast, high detecting endoscopists (baseline ADR ≥25%) achieved accelerated learning curves and sustained improvements in non-CADe procedures, suggesting that the exposure to AI support helped in improving already established detection strategies, for both adenomas and sessile serrated lesions. These findings were also supported by a recent nonrandomized controlled trial using CADe monitored by a second observer, which demonstrated a significant benefit of CADe for detectors with ADRs of 20%–40% [5]. The true long-term effect of CADe on endoscopy skills however was not evaluated in this study.

The favorable outcome reported by Okumura et al. contrasts with the conclusions of a recent multicenter observational study by Budzyń et al., which raised significant concerns about performance regression in non-CADe settings [6]. In their multicenter analysis, the authors reported a decrease in ADR after CADe withdrawal, especially among less experienced endoscopists, prompting the suggestion that AI may inhibit long-term skill development.

In contrast, Okumura et al. used a pragmatic single-center design and carefully categorized endoscopists by their baseline detection performance, not simply by procedural experience. This classification revealed an important nuance: while high detectors improved with and without CADe, low detectors (ADR <25%) did not show significant improved learning effects, even after extended CADe exposure. However, and even more crucially, their performance did not decline either. This results in a two-fold implication. First, CADe can serve as a learning adjunct for proficient endoscopists, with benefits that persist beyond the presence of the tool itself. Second, poorer performing endoscopists may require additional structured feedback or targeted educational strategies to fully benefit from CADe support. Therefore, the AI system alone may be insufficient to uniformly enhance skills across all users, analogous to how a GPS can chart optimal directions, but only if the driver has the ability to drive the car.

This differentiation echoes prior findings suggesting that baseline ADR, rather than total procedural volume, may be a more reliable indicator of an endoscopist’s potential to internalize and apply detection techniques [7]. The study of Okumura et al. supports this notion, indicating that personality traits such as vigilance and attention to detail, or technical experience, which may be more prevalent among high detectors, could influence the effectiveness of CADe for long-term skills improvement.

Another important strength of the authors’ study lies in its real-world longitudinal design. In contrast to randomized trials, which are susceptible to the potential Hawthorne effect where participants may alter their behavior owing to their awareness of being observed, the observational design of this study is more likely to accurately capture routine clinical practice. By incorporating CADe into daily practice and analyzing performance across both CADe and non-CADe cases, the authors mimic the actual variability of modern endoscopy units where AI tools are often selectively deployed.

Again, this underscores a few substantial implications. First, these findings reframe the current conversation around AI in colonoscopy, which is focused on the short-term efficacy and the fear of overdependence, to one that considers a sustained skill development and performance reinforcement in already skilled endoscopists but also keeps lower detectors from ultimate deskilling. Second, they offer reassurance to international guideline organizations and training institutions that CADe does not need to be seen as a jeopardy to clinical competence, provided its implementation is thoughtful and monitored, and on condition that the current observational findings can be reproduced in consecutive multicenter trials. Finally, the study results highlight the need for personalized AI-assisted training programs, particularly to address the learning needs of low detectors.

As the authors appropriately acknowledge, certain limitations persist. First and foremost, this was a single-center study, and the endoscopists included were not categorized as beginners, a group for whom skill retention is crucial. Second, despite the rigorous methodology, the study team also included CADe system developers, raising the potential for developer bias. Despite this, the balanced patient-matching, objective performance assessment, and granular stratification of detector status feed credibility to these results, although critical interpretation remains crucial.

In conclusion, Okumura et al. provide the first robust evidence that long-term CADe use does not deskill endoscopists and may in fact contribute to meaningful learning gains, albeit particularly among already proficient detectors. This study shifts the focus from caution for AI to calibration of AI, and stimulates thinking about how to implement CADe in a way that promotes individual growth for the endoscopist, but also safeguards quality.

With only few data available to date, the present study of Okumura et al. represents an important step toward answering a key question in the field of AI: as AI becomes more deeply embedded in endoscopy practice, the challenge is not whether we should use it, but how we can ensure it teaches us as much as it helps us.



Publication History

Article published online:
17 October 2025

© 2025. Thieme. All rights reserved.

Georg Thieme Verlag KG
Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany

Refers to:

Long-term impact of computer-aided adenoma detection: a prospective observational studyEndoscopy eFirst
DOI: 10.1055/a-2661-2624