Abstract:

Objective To construct a model for identifying venous crisis following skin flap transplantation and validate its efficacy，thereby providing objective evidence for the early and accurate recognition of venous crisis in skin flaps.Methods A prospective cohort study was conducted on patients who underwent flap transplantation at the microsurgery department of a tertiary-level hospital in Hunan Province from January 2022 to December 2024. Time-series images of the flaps during hospitalization were collected using Huawei M6 tablets to form the dataset. The Make Sense image processing software was employed to annotate regions of interest within the flaps and document their vascular status. Identification models for venous crisis after flap transplantation were established using recurrent neural networks，gated recurrent units，and long short-term memory networks，followed by internal validation. Model performance was evaluated using accuracy，precision，recall，F1 score，area under the receiver operating characteristic curve，Kolmogorov-Smirnov value for discrimination，calibration curve，and decision curve. The optimal model was selected.Results A total of 38，850 flap images from 661 patients were ultimately included. The accuracy rates of the recurrent neural network，gated recurrent unit，and long short-term memory network models were 0.948，0.949，and 0.984，respectively. Their precision rates were 0.603，0.449，and 0.776，while their recall rates were 0.711，0.700，and 0.922，respectively. Recall rates were 0.711，0.700，and 0.922，respectively. F1 scores were 0.607，0.529，and 0.811，respectively. The area under the receiver operating characteristic curve was 0.946，0.987，and 0.911，respectively，with discrimination rates of 0.885，0.963，and 0.965.Conclusion The flap venous crisis identification model constructed using flap time-series images and long short-term memory network method demonstrates excellent performance. It provides objective assessment criteria for nurses in monitoring flap perfusion，thereby facilitating early identification and intervention of venous crisis.

Key words: Flap Transplantation, Venous Crisis, Deep Learning, Medical Image, Predictive Model, Nursing Care

摘要：

目的 构建皮瓣移植术后静脉危象识别模型，并验证模型效果，为皮瓣静脉危象早期、准确识别提供客观依据。方法 采用便利抽样法，选取2022年1月—2024年12月在湖南省某三级甲等医院显微外科接受皮瓣移植手术的患者作为调查对象，进行前瞻性队列研究，使用平板电脑收集患者住院期间皮瓣的时间序列图像作为数据集，通过图像处理软件Make Sense对图像中感兴趣的皮瓣区域进行标注并注释血运状况。采用循环神经网络、门控循环单元、长短时记忆网络算法，建立皮瓣移植术后静脉危象风险识别模型并进行内部验证。通过准确度、精确度、召回率、F1值、受试者操作特征曲线下面积、区分度指标科尔莫戈洛夫-斯米尔诺夫值、校准曲线、决策曲线评价3种模型预测性能，选择最佳模型。结果 最终纳入来自661例患者的38 850张皮瓣图像，循环神经网络、门控循环单元、长短时记忆网络模型的准确度分别是0.948、0.949、0.984，精确度分别是0.603、0.449、0.776，召回率分别是0.711、0.700、0.922，F1值分别是0.607、0.529、0.811，受试者工作特征曲线下面积分别为0.946、0.987、0.911，区分度分别是0.885、0.963、0.965。结论 基于皮瓣时间序列图像和长短时记忆网络方法构建的皮瓣静脉危象识别模型具有良好性能，可为临床护士开展皮瓣血运监测提供客观的判断依据，有助于静脉危象的早期识别和干预。

关键词: 皮瓣移植, 静脉危象, 深度学习, 医学图像, 模型, 护理