Zhong Huang, Hou Cong, Huang Zhong, Chen Xinlian, Zou Yan, Zhang Han, Wang Tingyu, Wang Lan, Huang Xiangbing, Xiang Yongfeng, Zhong Ming, Hu Mingying, Xiong Dongmei, Wang Li, Zhang Yuanyuan, Luo Yan, Guan Yuting, Xia Mengyi, Liu Xiao, Yang Jinlin, Gan Tao, Wei Wei, Chen Honghan, Gong Hang
Department of Gastroenterology, Zigong First People's Hospital, Zigong, China.
Department of Gastroenterology and Hepatology, West China Hospital, Sichuan University, Chengdu, China.
Scand J Gastroenterol. 2025 Jan;60(1):116-121. doi: 10.1080/00365521.2024.2443520. Epub 2024 Dec 22.
High-quality bowel preparation is paramount for a successful colonoscopy. This study aimed to explore the effect of artificial intelligence-driven smartphone software on the quality of bowel preparation.
Firstly, we utilized 3305 valid liquid dung images collected mobile phones as training data. the most effective model was employed on mobile phones to evaluate the quality of bowel preparation. Secondly, From May 2023 to September 2023, colonoscopy patients were randomly assigned to two groups - the AI group ( = 116) and the control group ( = 116) - using a randomized, controlled, endoscopist-blinded method. We compared the two groups in terms of Boston Bowel Preparation Scale (BBPS) scores, polyp detection rate, adverse reaction rate, and factors related to bowel preparation quality. The primary endpoint was the percentage of patients who achieved a BBPS ≥6 among those who effectively utilized the smartphone software.
EfficientNetV2 exhibited the highest performance, with an accuracy of 87%, a sensitivity of 83%, and an AUC of 0.86. In the patient validation experiment, the AI group had higher BBPS scores than the control group (6.78 ± 1.41 vs. 5.35 ± 2.01, = 0.001) and showed an improvement in the detection rate (71.55% vs. 56.90%, = 0.020) for polyps. Multifactor logistic analysis indicated that compliance with enema solution usage rules (OR: 5.850, 95% confidence interval: 2.022-16.923), total water intake (OR: 1.001, 95% confidence interval: 1.001-1.002), and AI software reminders (OR: 2.316, 95% confidence interval: 1.096-4.893) were independently associated with BBPS scores ≥6.
Compared with traditional methods, the use of artificial intelligence combined with software to send reminders can lead to more accurate assessments of bowel preparation quality and an improved detection rate for polyps, thus demonstrating promising clinical value.
高质量的肠道准备对于成功进行结肠镜检查至关重要。本研究旨在探讨人工智能驱动的智能手机软件对肠道准备质量的影响。
首先，我们将收集到的3305张有效的液体粪便图像用作训练数据。在手机上使用最有效的模型来评估肠道准备质量。其次，从2023年5月至2023年9月，采用随机、对照、内镜医师盲法将结肠镜检查患者随机分为两组——人工智能组（n = 116）和对照组（n = 116）。我们比较了两组在波士顿肠道准备量表（BBPS）评分、息肉检出率、不良反应率以及与肠道准备质量相关的因素。主要终点是在有效使用智能手机软件的患者中达到BBPS≥6的患者百分比。
EfficientNetV2表现出最高的性能，准确率为87%，灵敏度为83%，曲线下面积为0.86。在患者验证实验中，人工智能组的BBPS评分高于对照组（6.78±1.41 vs. 5.35±2.01，P = 0.001），息肉检出率有所提高（71.55% vs. 56.90%，P = 0.020）。多因素逻辑分析表明，灌肠液使用规则的依从性（OR：5.850，95%置信区间：2.022 - 16.923）、总饮水量（OR：1.001，95%置信区间：1.001 - 1.002）以及人工智能软件提醒（OR：2.316，95%置信区间：1.096 - 4.893）与BBPS评分≥6独立相关。
与传统方法相比，使用人工智能结合软件发送提醒可更准确地评估肠道准备质量，并提高息肉检出率，从而显示出有前景的临床价值。
