AI辅助结肠镜对不同年资医师检出效应研究

Effect of Artificial Intelligence-Assisted Colonoscopy on Detection Outcomes among Endoscopists with Different Experience Levels

摘要: 目的：探讨人工智能(AI)辅助结肠镜对不同年资医师检出效应的影响。方法：采用单中心回顾性观察研究，纳入2021年9月至2022年9月内蒙古自治区人民医院住院结肠镜患者1093例。按是否使用AI分为传统组和AI组；按AI引入前医师累计结肠镜操作总数是否超过10,000例分为高、低年资组；以息肉检出率(PDR)和腺瘤检出率(ADR)为主要结局，进行总体比较、年资分层比较、AI应用初期与后期比较，并采用多因素logistic回归及双重差分全模型进行分析。结果：总体比较显示，AI组PDR高于传统组(61.6% vs 54.5%, P = 0.023)，ADR差异无统计学意义(43.6% vs 39.1%, P = 0.160)。分层比较显示，低年资层内AI组PDR高于传统组(64.6% vs 56.2%, P = 0.041)，高年资层内PDR和ADR差异均无统计学意义。分层多因素logistic回归显示，低年资层内AI与PDR升高独立相关(aOR = 1.63, 95% CI: 1.07~2.48, P = 0.024)，而其余结局相关性无统计学意义。AI应用初期与后期比较、亚组内前后多因素logistic回归及双重差分全模型均未见稳定统计学差异；但在低年资医师后期同期比较中，AI组PDR高于传统组(67.0% vs 54.8%, P = 0.038)。结论：AI辅助结肠镜可提高息肉检出率，且在低年资医师中更易体现检出优势；其对腺瘤检出率的促进作用尚需进一步验证。

Abstract: Objective: To investigate the effect of artificial intelligence (AI)-assisted colonoscopy on detection outcomes among endoscopists with different experience levels. Methods: This single-center retrospective observational study included 1,093 inpatients who underwent colonoscopy at Inner Mongolia Autonomous Region People’s Hospital from September 2021 to September 2022. Patients were divided into a conventional group and an AI-assisted group according to whether AI was used during colonoscopy. Endoscopists were further classified as senior or junior according to whether the cumulative number of colonoscopies performed before AI implementation exceeded 10,000. Polyp detection rate (PDR) and adenoma detection rate (ADR) were defined as the primary outcomes. Overall comparisons, stratified analyses by experience level, comparisons between the early and late phases after AI implementation, multivariable logistic regression, and difference-in-differences models were performed. Results: Overall, the AI-assisted group had a significantly higher PDR than the conventional group (61.6% vs 54.5%, P = 0.023), while the difference in ADR was not statistically significant (43.6% vs 39.1%, P = 0.160). Stratified analysis showed that among junior endoscopists, the AI-assisted group had a higher PDR than the conventional group (64.6% vs 56.2%, P = 0.041), whereas no significant differences in PDR or ADR were observed among senior endoscopists. Multivariable logistic regression demonstrated that AI use was independently associated with a higher PDR among junior endoscopists (aOR = 1.63, 95% CI: 1.07~2.48, P = 0.024), while no significant associations were found for the other outcomes. Comparisons between the early and late phases, subgroup multivariable analyses, and the overall difference-in-differences model did not show stable statistical differences; however, in the late period among junior endoscopists, the AI-assisted group showed a higher PDR than the conventional group (67.0% vs 54.8%, P = 0.038). Conclusion: AI-assisted colonoscopy improved polyp detection, and this advantage was more evident among junior endoscopists; its effect on adenoma detection still requires further validation.