多模态模型预测结肠息肉切除术后迟发性出血

doi:10.3969/j.issn.1674-8115.2026.04.011

上海交通大学学报（医学版） ›› 2026, Vol. 46 ›› Issue (4): 509-520.doi: 10.3969/j.issn.1674-8115.2026.04.011

• 论著 · 技术与方法 • 上一篇

多模态模型预测结肠息肉切除术后迟发性出血

诸梦琳¹, 刘骁², 徐晓丹¹, 王甘红³, 夏开建⁴, 陈健¹^,⁴()

^1.苏州大学附属常熟医院（江苏省常熟市第一人民医院）消化内科，常熟 215500
^2.江苏省常熟市尚湖中心医院消化内科，常熟 215500
^3.江苏省常熟市中医院（南京中医药大学常熟附属医院）消化内科，常熟 215500
^4.苏州大学附属常熟医院（江苏省常熟市第一人民医院）苏州市数据创新应用实验室，常熟 215500

收稿日期:2025-11-04 接受日期:2026-02-28 出版日期:2026-04-16 发布日期:2026-04-16
通讯作者: 陈健，副主任医师，硕士；电子信箱：szcs10132716@163.com。
基金资助:
苏州市临床重点病种诊疗技术专项(LCZX202334);苏州市科技攻关计划项目(SYW2025034);常熟市科技计划（社会发展）项目(CS202452);常熟市医学人工智能与大数据重点实验室能力提升项目(CYZ202301)

Prediction of delayed post-polypectomy bleeding using a multimodal model

Zhu Menglin¹, Liu Xiao², Xu Xiaodan¹, Wang Ganhong³, Xia Kaijian⁴, Chen Jian¹^,⁴()

^1.Department of Gastroenterology, Changshu Hospital Affiliated to Soochow University (Changshu No. 1 People's Hospital, Jiangsu Province), Changshu 215500, China
^2.Department of Gastroenterology, Changshu Shanghu Central Hospital, Jiangsu Province, Changshu 215500, China
^3.Department of Gastroenterology, Changshu Traditional Chinese Medicine Hospital (Changshu Hospital Affiliated to Nanjing University of Chinese Medicine), Jiangsu Province, Changshu 215500, China
^4.Suzhou Data Innovation Application Laboratory, Changshu Hospital Affiliated to Soochow University (Changshu No. 1 People's Hospital, Jiangsu Province), Changshu 215500, China

Received:2025-11-04 Accepted:2026-02-28 Online:2026-04-16 Published:2026-04-16
Contact: Chen Jian，E-mail：szcs10132716@163.com.
Supported by:
Special Project for Diagnosis and Treatment Technology of Clinical Key Diseases in Suzhou(LCZX202334);Suzhou Science and Technology Research Program Project(SYW2025034);Changshu Science and Technology Program (Social Development) Project(CS202452);Capacity Improvement Project of Changshu Key Laboratory of Medical Artificial Intelligence and Big Data(CYZ202301)

摘要/Abstract

摘要：

目的·构建一个多模态预测模型，将内镜创面图像与临床特征相结合，运用深度学习与机器学习技术，预测术后迟发性出血（delayed post-polypectomy bleeding，DPPB）的风险。方法·收集在2家医院行结肠息肉内镜下切除术患者的临床资料及术后内镜创面图像。研究共设计3个阶段：第一阶段，采用迁移学习训练的YOLOv11模型对创面图像进行自动识别与分割，提取感兴趣区域（region of interest，ROI）；第二阶段，以ROI图像为输入项，构建基于ResNet50架构的深度神经网络，结合术后是否发生DPPB进行监督学习，输出标准化的内镜影像学评分（endoscopic score，E-Score）；第三阶段，将E-Score与患者临床特征融合，利用LASSO回归进行特征选择，并将筛选后的变量输入多种机器学习模型训练，构建多模态DPPB预测模型。最终，基于Streamlit框架开发一款便于医护人员使用的网络应用工具（application，APP）。结果·共纳入2 782例接受结肠息肉切除术的患者，其中228例（8.20%）发生了DPPB。在测试集中，由XGBoost算法构建的多模态预测模型（PrismDPPB）表现最优，曲线下面积（area under the curve，AUC）达到0.831（95%CI 0.81~0.85），优于其他模型。此外，该模型在准确率（79.78%）、灵敏度（77.28%）、阳性预测值（81.32%）以及F1分数（79.25%）方面亦表现最佳。特征重要性分析显示，对模型预测贡献最大的6项变量依次为E-Score、息肉最大基底直径、性别、有无蒂、体质量指数以及息肉位置。模型解释性方面，使用SHAP图对预测结果进行了可视化解释。结论·融合创面影像组学特征与患者临床特征所构建的多模态预测模型及配套APP，在DPPB的风险评估中表现出良好的实用性与临床应用潜力。

关键词: 结肠息肉切除术, 术后迟发性出血, 网络应用工具, 深度学习, 机器学习, 多模态模型

Abstract:

Objective ·To develop a multimodal prediction model that integrates endoscopic wound images with clinical features, leveraging deep learning and machine learning techniques to predict the risk of delayed post-polypectomy bleeding (DPPB). Methods ·The clinical data and postoperative endoscopic wound images of patients who underwent endoscopic colorectal polypectomy at two hospitals were retrospectively collected. The study was designed in three stages. In the first stage, a YOLOv11 model trained via transfer learning was used to automatically detect and segment wound areas, extracting regions of interest (ROI). In the second stage, ROI images were input into a deep neural network based on the ResNet50 architecture. Supervised learning was performed using DPPB occurrence as the label, and a standardized endoscopic imaging score (E-Score) was generated. In the third stage, the E-Score was combined with clinical features. Feature selection was conducted using LASSO regression, and the selected variables were input into multiple machine learning algorithms to construct a multimodal DPPB prediction model. Finally, a web-based application was developed using the Streamlit framework to facilitate clinical use. Results ·A total of 2 782 patients who underwent colorectal polypectomy were included, among whom 228 (8.20%) developed DPPB. In the test set, the multimodal prediction model built using the XGBoost algorithm (PrismDPPB) achieved the best performance, with an area under the curve(AUC) of 0.831 (95%CI 0.81‒0.85), outperforming other models. It also showed superior performance in accuracy (79.78%), sensitivity (77.28%), positive predictive value (81.32%), and F1 score (79.25%). Feature importance analysis revealed that the six most contributive variables were E-Score, maximum polyp base diameter, sex, presence of a stalk, body mass index, and polyp location. For model interpretability, SHAP plots were used to visualize and explain the prediction results. Conclusion ·The multimodal prediction model developed by integrating image-based features with clinical characteristics, along with the accompanying web application, demonstrates strong practicality and clinical potential for DPPB risk assessment.

Key words: colorectal polypectomy, delayed post-polypectomy bleeding, web application, deep learning, machine learning, multimodal model

中图分类号:

R574.62

诸梦琳, 刘骁, 徐晓丹, 王甘红, 夏开建, 陈健. 多模态模型预测结肠息肉切除术后迟发性出血[J]. 上海交通大学学报（医学版）, 2026, 46(4): 509-520.

Zhu Menglin, Liu Xiao, Xu Xiaodan, Wang Ganhong, Xia Kaijian, Chen Jian. Prediction of delayed post-polypectomy bleeding using a multimodal model[J]. Journal of Shanghai Jiao Tong University (Medical Science), 2026, 46(4): 509-520.

图/表 12

图1 结肠息肉切除术后创面图像的分割流程示意图Note： A. Original endoscopic image. B. Manually annotated image. C. Segmentation results of model α with confidence box display. D. Extracted ROI image for subsequent radiomics analysis.

Fig 1 Schematic diagram of the segmentation workflow for post-polypectomy wound images

图2 特征筛选流程

Fig 2 Feature selection workflow

图3 网络APP开发流程

Fig 3 Web App development workflow

表1 训练集与测试集基线资料比较

Tab 1 Comparison of baseline data between the training set and the test set

Feature	Training set (n=1 947)	Test set (n=835)	t/Z/χ²	P value
Age/year	58.50±11.30	58.19±10.93	0.668	0.507
DBIL/(μmol·L^-1)	5.01 (1.73, 8.98)	4.79 (1.94, 8.14)	1.801	0.214
TBIL/(μmol·L^-1)	29.01 (10.97, 48.23)	26.47 (11.87, 46.33)	1.771	0.218
Cholesterol/(mmol·L^-1 )	7.88 (6.76, 8.98)	7.75 (6.68, 8.80)	1.959	0.047
AST/(U·L^-1 )	100.59 (78.61, 124.08)	101.89 (79.45, 122.13)	0.293	0.946
Platelet count/(10⁹·L^-1)	246.20 (186.34, 311.28)	249.63 (180.30, 311.17)	-0.023	0.982
BUN/(mg·dL^-1 )	6.33 (4.76, 7.86)	6.29 (4.85, 7.96)	-0.512	0.545
ALP/(U·L^-1 )	110.63 (81.65, 140.05)	109.15 (80.61, 140.26)	0.523	0.531
BMI/(kg·m^-2)	24.10 (22.10, 26.20)	24.00 (22.10, 26.20)	0.316	0.667
PT/s	11.73±0.82	11.70±0.86	0.804	0.415
INR	1.02 (0.98, 1.06)	1.02 (0.98, 1.06)	1.391	0.110
Albumin/(g·L^-1)	41.67 (38.33, 44.34)	41.59 (38.52, 44.49)	-0.711	0.712
Maximum basal diameter/cm	1.76 (1.33, 2.27)	1.69 (1.29, 2.19)	2.023	0.056
E-Score	0 (0, 0.02)	0 (0, 0.03)	-1.512	0.042
Smoking/n(%)			2.496	0.114
No	1 801 (92.5)	787 (94.3)
Yes	146 (7.5)	48 (5.7)
Alcohol consumption/n(%)			0.477	0.490
No	1 588 (81.6)	671 (80.4)
Yes	359 (18.4)	164 (19.6)
Hypertension/n(%)	949 (48.7)	420 (50.3)	0.507	0.477
No
Yes	998 (51.3)	415 (49.7)
Coronary heart disease/n(%)			0.342	0.559
No	344 (17.7)	156 (18.7)
Yes	1 603 (82.3)	679 (81.3)
Gender/n(%)			1.384	0.239
Female	1 026 (52.7)	461 (55.2)
Male	921 (47.3)	374 (44.8)
Diabetes mellitus/n(%)	1 614 (82.9)	675 (80.8)	1.560	0.212
No
Yes	333 (17.1)	160 (19.2)
Pedunculated status/n(%)			0.761	0.383
No	1 370 (70.4)	573 (68.6)
Yes	577 (29.6)	262 (31.4)
Antithrombotic drug use history/n(%)			0.010	0.982
No	1 847 (94.9)	793 (95.0)
Yes	100 (5.1)	42 (5.0)
Polyp location/n(%)			5.139	0.077
Left colon	769 (39.5)	334 (40.0)
Right colon	571 (29.3)	273 (32.7)
Bilateral colon	607 (31.2)	228 (27.3)
Number of polyps/n(%)			0.153	0.696
<3	1 551 (79.7)	659 (78.9)
≥3	396 (20.3)	176 (21.1)

图4 LASSO回归特征筛选过程Note： A. Coefficient path plot of LASSO regression. The horizontal axis represents Log λ, and the vertical axis represents the regression coefficient. As λ increases, the coefficients of some variables gradually shrink toward 0, and finally only the features with the strongest predictive power for the outcome are retained.B. Trend plot of binomial deviance in 5-fold cross-validation. As λ increases, the model deviance first decreases to a minimum and then gradually increases. The red dots represent the average deviance at each λ, and the gray lines represent the standard errors. The two dashed lines represent λ_min and λ_1se respectively.

Fig 4 Feature selection process using LASSO regression

图5 内镜影像组学流程Note： The input image first undergoes image segmentation to obtain the ROI. Subsequently, an end-to-end trained ResNet50 model is used to extract the E-Score from the segmented region. Then, the LASSO method is applied for feature selection. Based on the selected features, machine learning algorithms, including LR, RF, DCT, and XGBoost are used for modeling. Finally, model performance is evaluated using ROC curves, confusion matrices, and decision curve analysis. FN—false negative. FP—false positive. TN—true negative. TP—true positive.

Fig 5 Endoscopic radiomics workflow

图6 YOLO分割模型在训练集上的性能指标随训练进程的变化Note： A. Change in segmentation loss. B. Change in precision. C. Change in recall. D. Change in mAP50.

Fig 6 Changes in performance metrics of the YOLO segmentation model on the training set with training progress

图7 YOLOv11分割模型分割效果展示Note： A1‒A4. The original images of post-polypectomy wounds. B1‒B4. Segmentation results generated by the trained YOLO v11 segmentation model, where the red regions represent the irregular wound areas automatically identified by the model.

Fig 7 Visualization of segmentation performance of the YOLOv11 model

图8 5种模型在应用SMOTE过采样前后的AUC值对比Note： A. LR model. B. DCT model. C. XGBoost model. D. RF model. E. LNN model. F. Bar chart comparison.

Fig 8 Comparison of AUC values of five models before and after applying SMOTE oversampling

表2 5种模型的整体性能指标对比

Tab 2 Comparison of overall performance metrics among the five models

Model	Sensitivity/%	Specificity/%	Accuracy/%	PPV/%	NPV/%	F1 score/%	AUC (95%CI)
LR	71.80	76.14	73.97	75.03	73.00	73.38	0.809 (0.791‒0.833)
DCT	62.53	80.57	71.56	76.27	68.29	68.72	0.771 (0.752‒0.790)
RF	66.06	81.10	73.58	77.73	70.52	71.42	0.804 (0.781‒0.821)
XGBoost	77.28^①	82.27	79.78^①	81.32^①	78.39^①	79.25^①	0.831 (0.814‒0.852)
LNN	63.32	82.53^①	72.93	78.35	69.26	70.04	0.807 (0.790‒0.833)

图9 可供分享和便携使用的网络应用工具

Fig 9 Web application for sharing and portable use

图10 XGBoost 模型在测试集上的可视化解释

Fig 10 Visualization of the XGBoost model on the test set

参考文献 21

[1]	Bray F, Laversanne M, Sung H, et al. Global cancer statistics 2022: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries[J]. CA Cancer J Clin, 2024, 74(3): 229-263.
[2]	Lopes S R, Martins C, Santos I C, et al. Colorectal cancer screening: a review of current knowledge and progress in research[J]. World J Gastrointest Oncol, 2024, 16(4): 1119-1133.
[3]	Brenner H, Chang-Claude J, Seiler C M, et al. Protection from colorectal cancer after colonoscopy: a population-based, case-control study[J]. Ann Intern Med, 2011, 154(1): 22-30.
[4]	Shaukat A, Levin T R. Current and future colorectal cancer screening strategies[J]. Nat Rev Gastroenterol Hepatol, 2022, 19(8): 521-531.
[5]	Gupta S, Lieberman D, Anderson J C, et al. Recommendations for follow-up after colonoscopy and polypectomy: a consensus update by the US multi-society task force on colorectal cancer[J]. Gastrointest Endosc, 2020, 91(3): 463-485.e5.
[6]	Jaruvongvanich V, Prasitlumkum N, Assavapongpaiboon B, et al. Risk factors for delayed colonic post-polypectomy bleeding: a systematic review and meta-analysis[J]. Int J Colorectal Dis, 2017, 32(10): 1399-1406.
[7]	王禹毅, 卜志军, 李元晞, 等. 临床预测模型变量筛选方法及比较[J]. 现代中医临床, 2024, 31(2): 6-12.
	Wang Y Y, Bu Z J, Li Y X, et al. Variable selection methods and comparison in clinical prediction models[J]. Modern Clinical Chinese Medicine, 2024, 31(2): 6-12.
[8]	Cao J, Long S K, Liu H, et al. Constructing a prediction model for acute pancreatitis severity based on liquid neural network[J]. Sci Rep, 2025, 15(1): 16655.
[9]	Zhang R F, Yin M Y, Jiang A Q, et al. Application value of the automated machine learning model based on modified computed tomography severity index combined with serological indicators in the early prediction of severe acute pancreatitis[J]. J Clin Gastroenterol, 2024, 58(7): 692-701.
[10]	Liu L J, Zhang R F, Shi Y, et al. Automated machine learning for predicting liver metastasis in patients with gastrointestinal stromal tumor: a SEER-based analysis[J]. Sci Rep, 2024, 14(1): 12415.
[11]	Russell B C, Torralba A, Murphy K P, et al. LabelMe: a database and web-based tool for image annotation[J]. Int J Comput Vis, 2008, 77(1): 157-173.
[12]	Shin H C, Roth H R, Gao M C, et al. Deep convolutional neural networks for computer-aided detection: CNN architectures, dataset characteristics and transfer learning[J]. IEEE Trans Med Imaging, 2016, 35(5): 1285-1298.
[13]	Handelman G S, Kok H K, Chandra R V, et al. Peering into the black box of artificial intelligence: evaluation metrics of machine learning methods[J]. AJR Am J Roentgenol, 2019, 212(1): 38-43.
[14]	Nohara Y, Matsumoto K, Soejima H, et al. Explanation of machine learning models using shapley additive explanation and application for real data in hospital[J]. Comput Meth Programs Biomed, 2022, 214: 106584.
[15]	Fahmy A S, Csecs I, Arafati A, et al. An explainable machine learning approach reveals prognostic significance of right ventricular dysfunction in nonischemic cardiomyopathy[J]. JACC Cardiovasc Imaging, 2022, 15(5): 766-779.
[16]	Bendall O, James J, Pawlak K M, et al. Delayed bleeding after endoscopic resection of colorectal polyps: identifying high-risk patients[J]. Clin Exp Gastroenterol, 2021, 14: 477-492.
[17]	Zhang X Z, Jiang X X, Shi L. Risk factors for delayed colorectal postpolypectomy bleeding: a meta-analysis[J]. BMC Gastroenterol, 2024, 24(1): 162.
[18]	Albouys J, Montori Pina S, Boukechiche S, et al. Risk of delayed bleeding after colorectal endoscopic submucosal dissection: the Limoges Bleeding Score[J]. Endoscopy, 2024, 56(2): 110-118.
[19]	Lu Y, Zhou X Y, Chen H, et al. Establishment of a model for predicting delayed post-polypectomy bleeding: a real-world retrospective study[J]. Front Med, 2022, 9: 1035646.
[20]	王敏, 韦雪连, 李常伟, 等. 内镜下结直肠腺瘤性息肉的特征及经切除术后迟发出血的危险因素分析[J]. 胃肠病学和肝病学杂志, 2024, 33(6): 714-718.
	Wang M, Wei X L, Li C W, et al. Characteristics of colorectal adenomatous polyps under endoscopy and analysis of risk factors for delayed bleeding after resection[J]. Chinese Journal of Gastroenterology and Hepatology, 2024, 33(6): 714-718.
[21]	陈健, 王珍妮, 夏开建, 等. 基于深度学习的结直肠息肉内镜图像分割和分类方法比较[J]. 上海交通大学学报(医学版), 2024, 44(6): 762-772.
	Chen J, Wang Z N, Xia K J, et al. Comparative study on methods for colon polyp endoscopic image segmentation and classification based on deep learning[J]. Journal of Shanghai Jiao Tong University (Medical Science), 2024, 44(6): 762-772.

多模态模型预测结肠息肉切除术后迟发性出血

Prediction of delayed post-polypectomy bleeding using a multimodal model

RichHTML

PDF (PC)

可视化

摘要/Abstract

引用本文

使用本文

图/表 12

参考文献 21

相关文章 15

编辑推荐

Metrics

[1]	尹梓名, 王荣钦, 杨自逸, 刘颖斌, 陈涛, 束翌俊, 龚伟. 基于图神经网络的胆囊癌CT影像辅助诊断模型[J]. 上海交通大学学报（医学版）, 2025, 45(9): 1221-1231.
[2]	黄昕, 刘家辉, 叶敬文, 钱文莉, 许万星, 王琳. 基于机器学习的小细胞肺癌代谢分子诊断模型的建立和临床应用[J]. 上海交通大学学报（医学版）, 2025, 45(8): 1009-1016.
[3]	阮青青, 苏树智, 李延婷, 任渊, 戴勇, 乔增勇. 急性心肌梗死介入治疗并发症风险预测模型构建[J]. 上海交通大学学报（医学版）, 2025, 45(12): 1589-1597.
[4]	禹志远, 董海平, 高楠, 马柯. 背根神经节吗啡耐受核心基因筛选与机制研究：加权基因共表达网络分析和机器学习的转录组学整合策略[J]. 上海交通大学学报（医学版）, 2025, 45(10): 1308-1319.
[5]	吴其蓁, 刘启明, 柴烨子, 陶政宇, 王依楠, 郭欣宁, 姜萌, 卜军. 机器学习预测乳腺癌新辅助治疗后炎症代谢状态改变的模型评价[J]. 上海交通大学学报（医学版）, 2024, 44(9): 1169-1181.
[6]	许万星, 王琳, 郭巧梅, 王薛庆, 娄加陶. 多模态肺结节诊断模型的临床验证及应用价值探索[J]. 上海交通大学学报（医学版）, 2024, 44(8): 1030-1036.
[7]	陈健, 王珍妮, 夏开建, 王甘红, 刘罗杰, 徐晓丹. 基于深度学习的结直肠息肉内镜图像分割和分类方法比较[J]. 上海交通大学学报（医学版）, 2024, 44(6): 762-772.
[8]	刘洋, 吴梦怡, 胡尧, 亓坤, 王渝彬, 赵悦, 宋锦璘. 基于Transformer和扩散模型的头颅侧位片颈椎分割方法在正畸临床中的初步应用[J]. 上海交通大学学报（医学版）, 2024, 44(12): 1579-1586.
[9]	马奔, 赵成, 束翌俊, 董平. CT影像组学在胃肠道间质瘤中的应用进展[J]. 上海交通大学学报（医学版）, 2023, 43(7): 923-930.
[10]	赵珂珂, 蒋蓓蓓, 张璐, 王凌云, 张亚平, 解学乾. 超低剂量平扫CT深度学习图像重建评价肺部病灶的可行性[J]. 上海交通大学学报（医学版）, 2022, 42(8): 1062-1069.
[11]	林嘉希, 汪盛嘉, 赵鑫, 高欣, 殷民月, 朱锦舟. 基于深度卷积神经网络的Barrett食管内镜图片分类模型的建立[J]. 上海交通大学学报（医学版）, 2022, 42(5): 653-659.
[12]	陈立奇, 薛卓维, 吴氢凯. 基于磁共振成像的女性盆底器官三维数字模型重建的研究进展[J]. 上海交通大学学报（医学版）, 2022, 42(3): 381-386.
[13]	李欣, 范青. 机器学习在抑郁症患者面部特征研究中的应用进展[J]. 上海交通大学学报(医学版), 2022, 42(1): 124-129.
[14]	李小敏, 曲扬, 张少霆, 赵亮, 刘畅, 谢帅宁, 戴尅戎, 艾松涛. 人工智能技术在骨肌系统影像学方面的应用[J]. 上海交通大学学报(医学版), 2021, 41(2): 262-266.
[15]	陆言巧，沈兰，何奔. 人工智能在心血管疾病的辅助诊疗中的应用[J]. 上海交通大学学报（医学版）, 2020, 40(2): 259-.