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。

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.