腹腔镜胆囊切除术后住院延长预测研究进展

Research Progress on Prediction of Prolonged Length of Stay after Laparoscopic Cholecystectomy

摘要: 目的：综述腹腔镜胆囊切除术(laparoscopic cholecystectomy, LC)术后住院时间延长(prolonged length of stay, pLOS)的定义、影响因素与预测方法，为日间手术筛选、床位管理及围术期决策提供参考。方法：采用叙述性综述方法，检索PubMed、MEDLINE及相关英文数据库，检索时限截至2026年，纳入在线优先发表文献；检索范围涵盖pLOS、延迟出院、麻醉后恢复室(postanesthesia care unit, PACU)停留时间、日间手术失败、困难LC评分、加速康复外科(enhanced recovery after surgery, ERAS)路径、人工智能与机器学习(artificial intelligence/machine learning, AI/ML)预测模型及预测模型报告标准。文献按其涉及的临床终点分组解读，而非按发表年份归纳，以区分直接pLOS证据与邻近或间接证据；针对预测类研究，重点关注其终点定义、变量的临床可解释性、术前应用潜力及是否开展外部验证。结果：LC术后住院相关研究采用的终点异质性强，各自承载不同临床意义，不可合并分析。PACU停留时间主要反映麻醉恢复与复苏室资源占用；日间手术失败与术前筛选、手术排程及术后观察密切相关；术后 > 24 h延迟出院通常反映疼痛、呕吐、引流或并发症监测等早期恢复问题；以住院日≥3 d定义的pLOS更适用于择期LC后的风险分层。基于大型择期LC队列、以住院日≥3 d为终点构建的术前风险评分具有良好区分度。pLOS极少由单一因素导致，患者基础状态、年龄、美国麻醉医师协会(American Society of Anesthesiologists, ASA)分级、合并症负担、急性炎症、胆道疾病严重度、手术难度、术中事件、术后症状及出院标准，均在围术期不同阶段发挥作用。术前变量可用于医患沟通、日间手术准入评估与床位规划；术中变量有助于解释恢复复杂化的原因；术后症状或引流管留置常为延迟出院的直接诱因。困难LC评分可辅助解释延迟出院的手术背景，但不能替代pLOS预测模型，因其原始终点为手术难度而非住院时间。ERAS路径可缩短住院时间、改善恢复结局，包括减轻疼痛、降低术后恶心呕吐发生率、促进胃肠功能恢复；该获益源于多项围术期协同措施的累积效应，而非单一要素的作用，其获益幅度仍取决于本地路径的成熟度与出院标准。传统风险评分与回归模型因变量透明、具备临床可解释性，仍具有应用价值。AI/ML方法目前在LC中主要用于安全评估、解剖结构识别与工作流识别，直接预测pLOS的可靠证据仍有限。结论：LC术后pLOS缺乏统一定义，其阈值应结合本地出院标准校准，不可直接从单一研究队列外推，尤其在常规术后住院时间更长的医疗场景中。后续研究应统一终点定义，区分直接pLOS证据与邻近终点证据，在不同医疗体系与出院场景中开展外部验证，并按TRIPOD + AI声明规范报告模型。评价预测工具不应仅关注曲线下面积等区分度指标，更需关注其是否能优化日间手术筛选、围术期资源分配、ERAS管理强度及出院后随访。

Abstract: Objective: To review the definitions, influencing factors and prediction methods of prolonged length of stay (pLOS) after laparoscopic cholecystectomy (LC), so as to inform day-surgery selection, bed management and perioperative decision-making. Methods: This was a narrative review. PubMed, MEDLINE and related English-language databases were searched up to 2026, and online ahead-of-print articles were included. The search covered pLOS, delayed discharge, postanesthesia care unit (PACU) length of stay, day-case failure, difficult LC scores, enhanced recovery after surgery (ERAS) pathways, artificial intelligence and machine learning (AI/ML) prediction models, and reporting standards for prediction models. Studies were grouped and interpreted according to the clinical endpoint each study addressed, rather than by year of publication, so that direct pLOS evidence could be separated from adjacent or indirect evidence. For prediction studies, particular attention was paid to endpoint definition, clinical interpretability of the variables, potential for preoperative use, and whether external validation had been performed. Results: Studies of postoperative stay after LC use heterogeneous endpoints that carry different clinical meanings and should not be pooled. PACU length of stay mainly reflects anesthesia recovery and recovery-room resource use; day-case failure is more closely related to preoperative selection, scheduling and postoperative observation; delayed discharge beyond 24 hours usually reflects early recovery problems such as pain, vomiting, drainage or complication monitoring; and pLOS defined as a hospital stay of at least three days is more suitable for risk stratification after elective LC. A preoperative risk score derived in a large elective LC cohort, using a length of stay of at least three days as the endpoint, achieved good discrimination. pLOS is seldom caused by a single factor. Patient baseline status, age, American Society of Anesthesiologists (ASA) class, comorbidity burden, acute inflammation, biliary disease severity, operative difficulty, intraoperative events, postoperative symptoms and discharge criteria all contribute at different points along the perioperative pathway. Preoperative variables are useful for communication, day-surgery eligibility and bed planning; operative variables help explain why recovery becomes more complex; and postoperative symptoms or drainage often become the immediate reasons for delayed discharge. Difficult LC scores help explain the operative background of delayed discharge, but should not replace pLOS prediction models, because their original endpoint is operative difficulty rather than hospital stay. ERAS pathways shorten length of stay and improve recovery outcomes, including pain, postoperative nausea and vomiting, and gastrointestinal recovery; this benefit reflects the cumulative effect of several coordinated perioperative measures rather than any single component, and its magnitude still depends on local pathway maturity and discharge criteria. Conventional risk scores and regression-based models remain useful because their variables are transparent and clinically interpretable. AI/ML methods have been applied to LC mainly for safety assessment, anatomical recognition and workflow identification, whereas robust evidence for direct pLOS prediction is still limited. Conclusion: pLOS after LC lacks a unified definition, and its threshold should be calibrated to local discharge standards rather than extrapolated directly from a single cohort, especially where routine postoperative stay is longer. Future studies should standardize endpoint definitions, separate direct pLOS evidence from adjacent endpoints, carry out external validation across different health-care systems and discharge settings, and report models according to the TRIPOD + AI statement. A prediction tool should be judged not only by discrimination metrics such as the area under the curve, but also by whether it improves day-surgery selection, perioperative resource allocation, the intensity of ERAS management and post-discharge follow-up.