【作者】 张景伟；

【导师】 陈益强；

【作者基本信息】 郑州大学 ， 计算机科学与技术， 2025， 硕士

【摘要】 围手术期并发症是指患者在手术前后因手术刺激、麻醉药物或基础疾病等因素诱发的严重不良事件,严重影响患者术后恢复与生存率,是导致围手术期死亡和术后预后不良的主要因素之一。研究表明,尤其在老年患者及合并慢性病的高危人群中,并发症发生率显著升高。因此,围手术期对并发症风险的精准预测与早期干预,对于提高手术安全性、优化临床管理具有重要的现实意义。围术期涵盖患者术前准备、手术实施到术后恢复全过程,期间积累了大量异构医疗数据,例如术前结构化表格信息与术中动态血压监测数据。这些数据共同反映患者的风险状态,但由于结构差异显著,如何有效融合与建模始终是一个挑战。目前主流的风险评估方法(如修订心脏风险指数)多基于术前结构化变量,难以捕捉术中血压生理信号的动态变化,无法满足对实时预警的需求;而深度学习等方法虽然具备强大的建模能力,但普遍缺乏可解释性,难以获得临床医生信任。针对以上问题,本研究完成了以下工作:1.面向围术期并发症预测的多通道原型网络学习方法为克服现有方法在利用多通道血压预测并发症时,针对各自通道特点难以准确提取有用特征这一局限性,本文提出了多通道原型网络。该方法通过多通道原型匹配机制,得以充分提取收缩压、舒张压与平均动脉压的特征原型,既保留了各通道之间的结构独立性与生理差异性,又有效捕捉其关键时序模式。同时,设计了基于通道原型距离的自适应损失函数,在训练过程中动态优化通道间的判别性与类内一致性,从而提升模型的预测准确性与特征相关性。实验结果表明,该方法能准确预测围术期不良心血管事件的发生,并为临床提供可解释支持。2.面向异构数据融合的多模态原型网络建模方法针对术中时序血压数据与围术期结构化变量的异质模态融合以及可解释性不充分的问题,本研究提出了多模态原型网络。该网络通过多头协同注意力模块实现不同模态间的信息交互与对齐,有效建模血压时序特征与结构化临床变量之间的深层依赖关系。同时,引入多模态原型提取模块,从各模态中提取具有判别力的原型表示,通过原型间的融合增强特征表达能力,实现模态互补。此外,本研究设计了基于多模态原型聚类的损失函数,用以增强类内紧凑性与类间可分性,进一步提升原型表示的准确性与跨模态融合效果。实验结果表明,本方法不仅在性能上优于现有的多模态模型,而且通过对多模态原型特征进行可视化实现了良好的可解释性,能够为医生提供可信的决策支持。3.围术期术后并发症风险预测原型系统本研究基于上述两种方法,构建了围术期智能预警系统。该系统整合电子病历和术中监护数据,支持低延迟推理,并通过交互式可视化界面展示风险概率和关键特征。该系统能够辅助医生在短时间内识别并发现围术期并发症的发生,提高决策的可靠性。更多还原

【Abstract】 Perioperative complications refer to serious adverse events induced by surgical stimuli,anesthetic drugs,or underlying conditions before or after surgery.These com-plications significantly affect patients’postoperative recovery and survival,and are among the leading causes of perioperative mortality and poor postoperative progno-sis.Studies have shown that the incidence of complications is particularly high in el-derly patients and those with chronic diseases.Therefore,accurate risk prediction and early intervention for perioperative complications are of great practical significance in improving surgical safety and optimizing clinical management.The perioperative period encompasses the entire process from preoperative prepa-ration and surgical intervention to postoperative recovery,during which a large volume of heterogeneous medical data is generated,such as structured preoperative tabular in-formation and intraoperative dynamic blood pressure monitoring data.These data col-lectively reflect the patient’s risk status,but their structural differences pose significant challenges for effective integration and modeling.Current mainstream risk assessment methods(e.g.,Revised Cardiac Risk Index)are mostly based on preoperative structured variables and fail to capture dynamic changes in intraoperative physiological signals,thus limiting their utility in real-time warning.Although deep learning approaches of-fer strong modeling capabilities,they generally lack interpretability,making it difficult for clinicians to trust and adopt them.To address these issues,this study presents the following innovations:1.Multi-Channel Prototype Network for Perioperative Complication Predic-tionTo overcome the limitations of current methods in accurately extracting useful fea-tures from each individual blood pressure channel for complication prediction,we pro-pose a Multi-Channel Prototypical Network.This method leverages a multi-channel prototype matching mechanism to effectively extract prototypical temporal features from systolic,diastolic,and mean arterial pressure signals,preserving both structural in-dependence and physiological differences among channels while capturing critical tem-poral patterns.Additionally,an adaptive loss function based on channel-wise prototype distances is designed to dynamically optimize inter-channel discriminability and intra-class compactness during training.Experimental results demonstrate that this method can accurately predict major adverse cardiovascular events(MACE)during the periop-erative period and provide interpretable clinical support.2.Multi-Modal Prototype Network Modeling for Heterogeneous Data FusionTo address the challenges of heterogeneous modality fusion between intraopera-tive time-series blood pressure data and perioperative structured variables,as well as the lack of model interpretability,we propose a Multi-Modal Prototypical Network.This network utilizes a multi-head co-attention module to facilitate cross-modal inter-action and alignment,effectively modeling the deep dependencies between blood pres-sure sequences and structured clinical variables.Simultaneously,a prototype extraction module is introduced to capture discriminative representations from each modality and enhance feature expressiveness through prototype-level fusion,achieving complemen-tary integration.Furthermore,we design a prototype-based clustering loss function to improve the intra-class compactness and inter-class separability of the learned proto-types,thereby enhancing cross-modal fusion and interpretability.Experimental results show that the proposed method outperforms existing multimodal models in predictive performance and achieves strong interpretability via prototype visualization,offering trustworthy decision support for clinicians.3.Prototype System for Perioperative Postoperative Complication Risk Pre-dictionBased on the two proposed methods above,we developed an intelligent periopera-tive early warning system.The system integrates electronic health records and intraop-erative monitoring data,supports low-latency inference,and provides an interactive vi-sualization interface that displays risk probabilities and key contributing features.This system enables clinicians to identify and respond to perioperative complications in a timely manner,thereby improving the reliability of clinical decision-making.更多还原