危重症患者肠内喂养不耐受风险预测模型：一项系统评价。
全文（开放获取） · 中文译文
摘要 目的：系统评估危重症患者肠内喂养不耐受（EFI）风险预测模型，旨在为临床实践提供循证参考。 方法：系统检索PubMed、Web of Science、Cochrane Library、中国知网（CNKI）、万方数据库、中国生物医学文献数据库（CBM）及维普数据库，检索时限为建库至2025年11月20日。纳入开发或验证成人危重症患者肠内喂养不耐受风险预测模型的研究。采用CHARMS清单进行资料提取，并采用PROBAST工具评估所纳入研究的偏倚风险及适用性。 结果：共纳入25项研究，涉及25个肠内喂养不耐受风险预测模型。样本量为118至1,584例患者，EFI发生率为26.25%至61.3%。APACHE II评分、血糖、机械通气、年龄、白蛋白及腹内压是相对频繁报道的预测因子，但所有模型中未有一致的单一预测因子。 结论：现有的危重症患者肠内喂养不耐受风险预测模型显示出一定的预测价值；然而，方法学质量参差不齐，外部验证不足。未来研究应开展大规模、多中心、前瞻性研究，并采用标准化方法学程序来开发和验证预测模型，从而提高其临床适用性和可推广性。 系统综述注册：https://www.crd.york.ac.uk/PROSPERO/view/CRD420261281618，标识符（CRD420261281618）。 引言 危重症患者由于严重感染、创伤、休克及多器官功能障碍，常处于高代谢和高分解状态，导致营养不良的发生率显著增加（）。研究表明，营养不良与危重症患者的不良结局密切相关，包括感染率增加、机械通气时间和ICU住院时间延长，甚至死亡率增加（）。因此，早期、恰当的营养支持已成为危重症患者综合治疗中不可或缺的组成部分。 肠内营养（EN）遵循生理通路，有助于维持肠黏膜屏障功能、调节免疫反应并减少感染性并发症（）。因此，国内外多项指南推荐将其作为危重症患者的首选营养支持策略（）。国际和中国指南建议，对血流动力学稳定的患者尽早启动肠内营养以改善临床结局（）。然而，在临床实践中，喂养不耐受（EFI）在危重症患者中较为常见，文献报道的发生率在20%至60%之间，各研究间差异较大（,）。喂养不耐受通常表现为胃残余量增加、腹胀、呕吐、腹泻、反流和误吸，可能导致肠内营养中断、实际能量摄入不足，以及吸入性肺炎、感染和死亡风险增加（,）。 为了降低肠内喂养不耐受的风险，研究人员尝试建立风险预测模型来估计危重患者肠内喂养不耐受的概率（）。这些模型通常基于人口统计学特征、疾病严重程度评分、血流动力学状态、药物使用和胃肠功能指标，并使用诸如逻辑回归等统计方法构建。模型性能通常使用区分度、校准度和验证指标进行评估（）。一项多中心整群随机对照试验已被设计用于评估预测模型指导的喂养方案是否能改善肠内营养管理和患者结局；然而，临床有效性数据尚不可用（）。 然而，目前关于危重患者肠内喂养不耐受风险预测模型的研究在研究对象、结局定义、预测因子选择和模型验证方法方面存在显著异质性。因此，这些模型的预测性能和临床适用性需要系统评估。对现有研究的全面系统性评价是必要的。 基于这一背景，本研究旨在系统总结危重症患者肠内喂养不耐受风险预测模型的研究现状，分析模型构建方法、预测因素和预测性能，评估偏倚风险和临床适用性，并为临床实践和未来研究提供循证指导。 材料与方法 文献纳入与排除标准 纳入标准：（1）接受肠内营养的成年危重症患者；（2）研究内容为危重症患者肠内喂养不耐受风险预测模型的构建或验证；（3）研究设计包括队列研究、横断面研究和病例对照研究；（4）以中文或英文发表的文章。排除标准：（1）仅分析危险因素而未建立预测模型的研究；（2）数据不完整或无法获取全文；（3）重复发表的文献；（4）会议摘要；（5）基于系统评价或荟萃分析构建模型的研究。本系统评价已在PROSPERO注册（注册号：CRD420261281618）。 文献检索策略 系统检索了PubMed、Web of Science、Cochrane Library、中国知网（CNKI）、万方数据库、中国生物医学文献数据库（CBM）和维普数据库，以寻找关于危重症患者肠内喂养不耐受风险预测模型开发和验证的研究。检索时间范围从建库至2025年11月20日，文献语言限制为中文或英文。采用主题词和自由词相结合，并辅以手工检索参考文献列表。检索词包括：危重疾病、重症监护病房、ICU、重症监护、肠内营养、管饲、胃内喂养、幽门后喂养、胃肠道不耐受、胃肠疾病、喂养不耐受、风险预测模型、风险评估、Predict*、预测模型。 本系统评价采用CHARMS清单推荐的PICOTS框架来确定评价目标、检索策略和纳入/排除标准： P（人群）：收入重症监护病房并接受肠内营养的成年危重症患者。 I（指标预测模型）：用于估计肠内喂养不耐受风险或概率的预测模型。 C（比较模型）：无竞争模型。 O（结局）：导致肠内营养中断或喂养不足的胃肠道症状或事件。 T（时机）：在肠内营养开始前或开始时，结局在ICU住院期间评估。 S（场景）：医院环境中的重症监护病房。 文献筛选与数据提取 使用 Note Express 软件去除重复记录。两名接受过循证方法学培训的研究者独立筛选标题和摘要，随后根据纳入和排除标准进行全文审查。分歧通过讨论或咨询第三名研究者解决。 数据提取以 CHARMS 清单为指导，该清单由 Altman 和 Moons 于 2014 年制定（）。提取的信息包括第一作者、研究设计、发表年份、数据来源时期、建模方法、验证方法、预测因素及模型性能（）。数据提取由两名研究者独立进行，分歧由第三名研究者解决。对于连续且动态变化的预测因素，我们进一步提取原始研究是否报告了测量时机、建模形式、截断值、阈值选择方法，或非线性/时间依赖性效应的评估。我们还提取了 EFI 定义的详细组成，包括是否使用了 GRV、GRV 阈值、测量时机、GRV 是作为单次测量还是 24 小时累积量进行评估，以及是否存在其他临床表现，如呕吐、反流、腹泻、腹胀、误吸、消化道出血、肠鸣音改变和未达到目标能量摄入。此外，我们提取了纳入研究是否评估了模型指导的临床干预或实施效果。 偏倚风险与适用性评估 使用Wolff等人于2019年（）开发的预测模型偏倚风险评估工具（PROBAST）评估所纳入研究的偏倚风险和应用性。该工具评估四个领域的偏倚风险，包括参与者、预测因子、结局和统计分析，并评估三个领域的适用性，即参与者、预测因子和结局。PROBAST工具中的每个信号问题均以“是”、“可能是”、“可能否”、“否”或“无信息”作答，领域层面的偏倚风险判断根据回答分为低、高或不明确。如果一个或多个领域被判断为高偏倚风险，则该研究的总体偏倚风险被视为高。两名研究人员使用PROBAST工具（，）独立对所有纳入研究进行了偏倚风险和应用性评估。若意见不一致，则咨询第三名研究人员，并通过讨论达成共识。评估结果以表格形式汇总和呈现，包括每项纳入研究的偏倚风险和应用性的领域特定判断及总体判断，详细结果见。 结果 文献检索过程 共识别出3,026条记录。去除重复项并经标题和摘要筛选后，剩余31篇文章。经全文审查后，纳入25项研究（）。PRISMA流程图如图所示。 PRISMA流程图。 纳入研究的基本特征 本研究共纳入25项研究（），其中24项研究（，）由中国学者发表，1项研究（）由以色列学者发表。在发表时间方面，24项研究发表于过去5年内。在研究人群方面，所有研究均涉及接受肠内营养支持的成年重症患者，包括脓毒症患者、重症卒中患者、重症急性胰腺炎患者、神经外科重症患者、接受俯卧位通气的患者以及接受无创通气的患者。在研究设计方面，15项研究（，，，，，，，）为回顾性队列研究，6项研究（，，，，）为前瞻性队列研究，4项研究（，，，）为病例对照研究。19项研究（，，，，，）为单中心研究，6项研究（，，，，，）为多中心研究。纳入研究的样本量范围为118例至1,584例患者。肠内喂养不耐受的结果定义主要包括胃残余量增加、呕吐、腹胀、腹泻、误吸、消化道出血，以及因不耐受导致的肠内营养中断或能量摄入不足。纳入研究的基本特征见。 各纳入研究对肠内喂养不耐受（EFI）的定义不尽相同。胃残余量（GRV）是最常报告的客观组成部分，尽管报告的阈值和测量时间窗差异很大。两项研究将肠内营养启动后 4 h 测得的单次 GRV ≥ 200 mL 定义为 EFI (, )，而 1 项研究使用 6 h 内单次 GRV > 200 mL 且 6 h 内累积 GRV ≥ 250 mL ()，另一项采用 6 h 后 GRV > 250 mL 的阈值 ()。其他研究使用 24 h 累积阈值，包括 GRV ≥ 200 mL/24 h ()、GRV > 250 mL/24 h () 和 GRV ≥ 500 mL/24 h (, )。一些研究报告了胃潴留或 GRV 升高，但未指定定量阈值 (, , , , )。除 GRV 外，EFI 的定义通常还包括呕吐、反流、腹泻、腹胀、误吸、消化道出血、肠鸣音减弱或消失、肠内营养中断或暂停，或 72 h 内未达到目标能量摄入 (, , , , , , , )。 根据所报告的组成成分，EFI 定义被分为四类：单时间点 GRV 定义；累积或 24 h GRV 定义；将 GRV 与胃肠道症状和/或能量摄入不足相结合的复合定义；以及无量化 GRV 阈值的基于症状的定义。大多数研究报告了整体预测模型的敏感度和特异度，但未报告各 EFI 组分的性能指标。 在纳入的25项研究中（），肠内喂养不耐受的发生率范围为26.25%至61.3%。所有25项研究均使用逻辑回归，单独使用或与其他建模方法结合使用。五项研究（、、、、）采用了随机森林算法，三项研究（、、）使用了支持向量机，两项研究（、）使用了梯度提升树，一项研究（）使用了深度学习，一项研究（）使用了XGBoost，两项研究（、）使用了k近邻算法，一项研究（）使用了AdaBoost，两项研究（、）使用了决策树，一项研究（）使用了朴素贝叶斯算法。 所有25项研究均进行了内部验证，九项研究（、、、）进行了外部验证。十七项研究（、、、、、、）使用了自举重采样法，六项研究（、、、、、）使用了随机分割验证，一项研究（）使用了十折交叉验证，一项研究（）使用了五折交叉验证。二十四项研究（）报告了AUC值，其范围为0.70至0.99。在比较多种算法的研究中，比较主要基于AUC等区分度指标。校准和外部验证的报道较少。详细信息见。 纳入模型中的预测因素 所有25项纳入研究（）均报告了其预测模型的呈现格式。20项研究呈现了列线图，10项研究（, , , , , , , , , ）提供了回归方程，5项研究（, , , , ）呈现了机器学习模型。APACHE II评分、年龄、白蛋白、血糖、机械通气和腹内压是最常纳入的预测因子。血糖和腹内压是最常报告的连续预测因子之一。测量时机、建模形式、截断值、截断值选择方法、非线性关联、时间依赖性变化和重复测量的报告在各研究间有所不同。各个模型包含的呈现格式和预测因子显示在。 风险预测模型的模型呈现与预测因子。 APACHE II，急性生理与慢性健康评估-II，NRS2002，营养风险筛查2002，AGI，急性胃肠损伤，GCS，格拉斯哥昏迷评分，mNutric，改良危重症营养风险，CRRT，连续性肾脏替代治疗，SOFA，序贯器官衰竭评估。 纳入研究的方法学质量评估 研究人群领域 在纳入的研究中，15项研究（, , , , , , , ）为回顾性队列研究，19项研究（, , , , , ）为单中心研究。这些研究被判定为具有高偏倚风险。 预测因子领域 所有25项纳入研究（）对预测因子在不同研究人群中采用了一致的定义和评估方法，并明确规定了评价标准。大多数纳入研究清楚描述了预测因子的定义和测量方法，且预测因子通常基于常规可获得的临床变量。因此，预测因子领域被认为具有相对较低的适用性担忧。 结局领域 所有25项研究（）在结局领域被评定为具有相对较高的偏倚风险。EFI的操作性定义在各研究中有所不同。报告的结局组分包括不同的GRV阈值、呕吐、反流、腹泻、腹胀、误吸、消化道出血、肠鸣音改变、肠内营养中断或暂停，以及未达到目标能量摄入。 统计分析领域 所有25项研究（）在统计分析领域被评定为具有较高的偏倚风险。6项研究（, , , , , ）在模型开发或验证期间样本量不足。19项研究（, , , , , ）为单中心研究，6项研究（, , , , , ）采用随机分割验证，3项研究（, , ）未报告校准，9项研究（, , , ）进行了外部验证，4项研究（, , , ）未明确报告处理缺失数据的方法。 适用性评估 在受试者、预测因子和结局领域评估了适用性。根据PROBAST，所有25项研究（）在这三个领域的适用性均被评为低顾虑。偏倚风险和适用性的详细评估见。 预测模型的临床影响和干预效果 所纳入研究均未直接评估由所开发预测模型指导的干预措施的效果。尽管研究报告了模型性能指标，如AUC、灵敏度、特异度、校准和验证结果，但没有任何研究评估在临床实践中应用该模型是否改善了肠内营养实施或患者结局。具体而言，所纳入研究没有证据表明，模型指导的管理降低了肠内喂养不耐受的发生率、提高了目标能量达标率、减少了喂养中断、误吸、ICU住院时间或死亡率。 讨论 常见报告的预测因子及模型间的异质性 本系统评价的结果显示，纳入的模型在候选预测因子选择和最终预测因子保留方面均存在显著异质性。没有哪个单一的预测因子在所有纳入研究中都被一致识别。APACHE II 评分、血糖、机械通气、年龄、白蛋白和腹内压是相对较常报告的预测因子，但每个仅出现在一部分模型中。因此，这些预测因子应被解读为反复出现的信号或产生假说的因素，而非肠内喂养不耐受的确定性或普遍适用的预测因子。研究人群、EFI定义、候选预测因子集合、变量选择方法和验证策略的差异，均可能导致了各模型保留预测因子的异质性。 有11项研究 (, , , , , ) 将APACHE II评分确定为肠内喂养不耐受的预测因子。APACHE II是生理紊乱和疾病严重程度的整体测量指标，而非胃肠特异性指标。关于胃肠动力障碍和喂养不耐受的综述，需参考相关文献。
全文（开放获取） · 英文原文
【Abstract】 OBJECTIVE: To systematically evaluate existing risk prediction models for enteral feeding intolerance (EFI) in critically ill patients, with the aim of providing evidence-based references for clinical practice. METHODS: PubMed, Web of Science, the Cochrane Library, China National Knowledge Infrastructure (CNKI), Wanfang Database, China Biology Medicine Disc (CBM), and VIP Database were systematically searched from database inception to November 20, 2025. Studies that developed or validated risk prediction models for enteral feeding intolerance in adult critically ill patients were included. Data extraction was conducted using the CHARMS checklist, and the risk of bias and applicability of the included studies were assessed using the PROBAST tool. RESULTS: A total of 25 studies were included, involving 25 risk prediction models for enteral feeding intolerance. Sample sizes ranged from 118 to 1,584 patients, and the incidence of EFI ranged from 26.25 to 61.3%. APACHE II score, blood glucose, mechanical ventilation, age, albumin, and intra-abdominal pressure were relatively frequently reported predictors, but no single predictor was consistently identified across all models. CONCLUSION: Existing risk prediction models for enteral feeding intolerance in critically ill patients demonstrate certain predictive value; however, the methodological quality varies considerably, and external validation is insufficient. Future studies should conduct large-scale, multicenter, prospective research and adopt standardized methodological procedures to develop and validate prediction models, thereby improving their clinical applicability and generalizability. SYSTEMATIC REVIEW REGISTRATION: https://www.crd.york.ac.uk/PROSPERO/view/CRD420261281618, identifier (CRD420261281618). 【Introduction】 Critically ill patients, due to severe infections, trauma, shock, and multiple organ dysfunction, are often in a hypermetabolic and hypercatabolic state, resulting in a significantly increased incidence of malnutrition (). Studies have shown that malnutrition is closely associated with adverse outcomes in critically ill patients, including increased infection rates, prolonged duration of mechanical ventilation and ICU stay, and even increased mortality (). Therefore, early and appropriate nutritional support has become an essential component of comprehensive treatment in critically ill patients. Enteral nutrition (EN), which follows physiological pathways, helps maintain intestinal mucosal barrier function, modulate immune responses, and reduce infectious complications (). Consequently, it is recommended by multiple national and international guidelines as the preferred nutritional support strategy for critically ill patients (). International and Chinese guidelines recommend initiating enteral nutrition as early as possible in hemodynamically stable patients to improve clinical outcomes (). However, in clinical practice, enteral feeding intolerance (EFI) is common among critically ill patients, with reported incidence rates ranging from 20 to 60%, varying widely across studies (, ). Feeding intolerance typically manifests as increased gastric residual volume, abdominal distension, vomiting, diarrhea, reflux, and aspiration, which may lead to interruption of enteral nutrition, insufficient actual energy intake, and increased risks of aspiration pneumonia, infection, and mortality (, ). To reduce the risk of enteral feeding intolerance, researchers have attempted to develop risk prediction models to estimate the probability of feeding intolerance in critically ill patients (). These models are usually based on demographic characteristics, disease severity scores, hemodynamic status, medication use, and gastrointestinal function indicators, and are constructed using statistical methods such as logistic regression. Model performance is commonly evaluated using discrimination, calibration, and validation measures (). A multicenter cluster-randomized controlled trial has been designed to evaluate whether a prediction-model-guided feeding protocol can improve enteral nutrition management and patient outcomes; however, clinical effectiveness data are not yet available (). However, current studies on risk prediction models for enteral feeding intolerance in critically ill patients show substantial heterogeneity in study populations, outcome definitions, predictor selection, and model validation methods. Therefore, the predictive performance and clinical applicability of these models require systematic evaluation. A comprehensive systematic review of existing studies is necessary. Based on this background, the present study aims to systematically summarize the current status of research on risk prediction models for enteral feeding intolerance in critically ill patients, analyze model development methods, predictive factors, and predictive performance, evaluate the risk of bias and clinical applicability, and provide evidence-based guidance for clinical practice and future research. 【Materials and methods】 【Literature inclusion and exclusion criteria】 Inclusion criteria: (1) Adult critically ill patients receiving enteral nutrition; (2) Studies focused on the development or validation of risk prediction models for enteral feeding intolerance in critically ill patients; (3) Study designs including cohort studies, cross-sectional studies, and case–control studies; (4) Articles published in Chinese or English. Exclusion criteria: (1) Studies analyzing only risk factors without establishing a prediction model; (2) Incomplete data or inaccessible full texts; (3) Duplicate publications; (4) Conference abstracts; (5) Studies constructing models based on systematic reviews or meta-analyses. This systematic review has been registered with PROSPERO (registration ID: CRD420261281618). 【Literature search strategy】 PubMed, Web of Science, the Cochrane Library, CNKI, Wanfang Database, China Biology Medicine Disc (CBM), and VIP Database were systematically searched for studies on the development and validation of risk prediction models for enteral feeding intolerance in critically ill patients. The search period covered database inception to November 20, 2025, with publications limited to Chinese or English. Both subject headings and free-text terms were used, supplemented by manual searching of reference lists. Search terms included: Critical Illness, Intensive Care Units, ICU, Critical Care, Enteral Nutrition, tube feeding, gastric feeding, postpyloric feeding, Gastrointestinal Intolerance, Gastrointestinal Diseases, Feeding Intolerance, Risk Prediction Models, Risk Assessment, Predict*, Prediction Model. This systematic review adopted the PICOTS framework recommended by the CHARMS checklist to define evaluation objectives, search strategies, and inclusion/exclusion criteria:P (Population): Adult critically ill patients admitted to intensive care units and receiving enteral nutrition. I (Index prognostic model): Prediction models developed to estimate the risk or probability of enteral feeding intolerance. C (Comparative model): No competing model. O (Outcome): Gastrointestinal symptoms or events leading to interruption or inadequate delivery of enteral nutrition. T (Timing): Before or at initiation of enteral nutrition, with outcomes assessed during ICU stay. S (Setting): Intensive care units in hospital settings. 【Literature screening and data extraction】 Note Express software was used to remove duplicate records. Two researchers trained in evidence-based methodology independently screened titles and abstracts, followed by full-text review based on inclusion and exclusion criteria. Discrepancies were resolved through discussion or consultation with a third researcher. Data extraction was guided by the CHARMS checklist, developed by Altman and Moons in 2014 (). Extracted information included first author, study design, publication year, data source period, modeling method, validation method, predictive factors, and model performance (). Data extraction was conducted independently by two researchers, with disagreements resolved by a third researcher. For continuous and dynamically changing predictors, we further extracted whether the original studies reported measurement timing, modeling form, cut-off values, threshold-selection methods, or assessment of nonlinear/time-dependent effects. We also extracted detailed components of EFI definitions, including whether GRV was used, GRV thresholds, measurement timing, whether GRV was assessed as a single measurement or cumulative 24-h volume, and the presence of other clinical manifestations such as vomiting, reflux, diarrhea, abdominal distension, aspiration, gastrointestinal bleeding, bowel sound changes, and failure to achieve target energy intake. In addition, we extracted whether the included studies evaluated model-guided clinical interventions or implementation effects. 【Risk of bias and applicability assessment】 The risk of bias and applicability of the included studies were assessed using the Prediction model Risk Of Bias ASsessment Tool (PROBAST) developed by Wolff et al. in 2019 (). The tool evaluates the risk of bias across four domains, including participants, predictors, outcomes, and statistical analysis, and assesses applicability across three domains, namely participants, predictors, and outcomes. Each signaling question in the PROBAST tool is answered as “yes,” “probably yes,” “probably no,” “no,” or “no information,” and domain-level judgments of risk of bias are classified as low, high, or unclear based on the responses. If one or more domains are judged as having a high risk of bias, the overall risk of bias of the study is considered high. Two researchers independently conducted the risk of bias and applicability assessment for all included studies using the PROBAST tool (, ). In cases of disagreement, a third researcher was consulted, and consensus was reached through discussion. The assessment results were summarized and presented in tabular form, including both domain-specific judgments and overall judgments of risk of bias and applicability for each included study, and the detailed results are shown in . 【Results】 【Literature search process】 A total of 3,026 records were identified. After removing duplicates and screening titles and abstracts, 31 articles remained. Following full-text review, 25 studies were included (). The PRISMA flow diagram is shown in . PRISMA flow diagram. 【Basic characteristics of included studies】 This study included a total of 25 studies (), of which 24 studies (, ) were published by Chinese scholars and 1 study () was published by Israeli scholars. In terms of publication time, 24 studies were published within the past 5 years. With respect to study populations, all studies involved adult critically ill patients receiving enteral nutrition support, including patients with sepsis, severe stroke, severe acute pancreatitis, neurosurgical critically ill patients, patients receiving prone position ventilation, and patients receiving noninvasive ventilation. Regarding study design, 15 studies (, , , , , , , ) were retrospective cohort studies, 6 studies (, , , ) were prospective cohort studies, and 4 studies (, , , ) were case–control studies. Nineteen studies (, , , , , ) were single-center studies, and 6 studies (, , , , , ) were multicenter studies. The sample sizes of the included studies ranged from 118 to 1,584 patients. Outcome definitions of enteral feeding intolerance mainly included increased gastric residual volume, vomiting, abdominal distension, diarrhea, aspiration, gastrointestinal bleeding, and interruption of enteral nutrition or insufficient energy intake due to intolerance. The basic characteristics of included studies are shown in . 【Variations in EFI definitions and GRV measurement】 The definitions of enteral feeding intolerance varied across the included studies. GRV was the most commonly reported objective component, although the reported thresholds and measurement windows differed considerably. Two studies defined EFI using a single GRV ≥ 200 mL measured 4 h after enteral nutrition initiation (, ), whereas 1 study used a single GRV > 200 mL within 6 h together with a cumulative GRV ≥ 250 mL within 6 h (), and another adopted a threshold of GRV > 250 mL after 6 h (). Other studies used cumulative 24-h thresholds, including GRV ≥ 200 mL/24 h (), GRV > 250 mL/24 h (), and GRV ≥ 500 mL/24 h (, ). Several studies reported gastric retention or elevated GRV without specifying a quantitative threshold (, , , , ). In addition to GRV, EFI definitions commonly incorporated vomiting, reflux, diarrhea, abdominal distension, aspiration, gastrointestinal bleeding, decreased or absent bowel sounds, interruption or suspension of enteral nutrition, or failure to achieve target energy intake within 72 h (, , , , , , , ). Based on the reported components, EFI definitions were classified into four categories: single-time-point GRV definitions; cumulative or 24-h GRV definitions; composite definitions combining GRV with gastrointestinal symptoms and/or inadequate energy intake; and symptom-based definitions without a quantified GRV threshold. Most studies reported sensitivity and specificity for the overall prediction model but did not report performance measures for individual EFI components. 【Methods for model construction and predictive performance】 Among the 25 included studies (), the incidence of enteral feeding intolerance ranged from 26.25 to 61.3%. All 25 studies used logistic regression, either alone or in combination with other modeling approaches. Five studies (, , , , ) employed random forest algorithms, three studies (, , ) used support vector machines, two studies (, ) used gradient boosting trees, one study () used deep learning, one study () used XGBoost, two studies (, ) used k-nearest neighbor algorithms, one study () used AdaBoost, two studies (, ) used decision trees, and one study () used a naive Bayes algorithm. All 25 studies performed internal validation, and nine studies (, , , ) conducted external validation. Seventeen studies (, , , , , , ) used bootstrap resampling, six studies (, , , , , ) used random split validation, one study () used 10-fold cross-validation, and one study () used five-fold cross-validation. Twenty-four studies (, ) reported AUC values, which ranged from 0.70 to 0.99. In studies comparing multiple algorithms, comparisons were primarily based on discrimination measures such as AUC. Calibration and external validation were less frequently reported. Detailed information is presented in . 【Predictive factors included in the models】 All 25 included studies () reported the presentation format of their prediction models. Twenty studies presented nomograms, ten studies (, , , , , , , , , ) provided regression equations, and five studies (, , , , ) presented machine learning models. APACHE II score, age, albumin, blood glucose, mechanical ventilation, and intra-abdominal pressure were the most frequently included predictors. Blood glucose and intra-abdominal pressure were among the most frequently reported continuous predictors. Reporting of measurement timing, modeling form, cut-off values, cut-off selection methods, nonlinear associations, time-dependent changes, and repeated measurements varied across studies. The presentation formats and predictors included in individual models are shown in . Model presentation and predictive factors of risk prediction models. APACHE II, Acute Physiology and Chronic Health Evaluation-II, NRS2002, Nutrition Risk Screening 2002, AGI, Acute Gastrointestinal Injury, GCS, Glasgow Coma Scale, mNutric, modified nutrition risk in the critically ill, CRRT, continuous renal replacement therapy, SOFA, Sequential Organ Failure Assessment. 【The methodological quality assessment of included studies】 【Study population domain】 Among the included studies, 15 studies (, , , , , , , ) were retrospective cohort studies, and 19 studies (, , , , , ) were single-center studies. These studies were judged to have a high risk of bias. 【Predictor domain】 All 25 included studies () used consistent definitions and assessment methods for predictors across study populations and clearly specified evaluation criteria. Most included studies clearly described predictor definitions and measurement methods, and predictors were generally based on routinely available clinical variables. Therefore, the predictor domain was considered to have relatively low applicability concerns. 【Outcome domain】 All 25 studies () were rated as having a relatively high risk of bias in the outcome domain. The operational definitions of EFI varied across studies. Reported outcome components included different GRV thresholds, vomiting, reflux, diarrhea, abdominal distension, aspiration, gastrointestinal bleeding, bowel sound changes, interruption or suspension of enteral nutrition, and failure to achieve target energy intake. 【Statistical analysis domain】 All 25 studies () were rated as having a high risk of bias in the statistical analysis domain. Six studies (, , , , , ) had insufficient sample sizes during model development or validation. Nineteen studies (, , , , , ) were single-center studies, six studies (, , , , , ) used random split validation, three studies (, , ) did not report calibration, nine studies (, , , ) conducted external validation, and four studies (, , , ) did not clearly report methods for handling missing data. 【Applicability assessment】 Applicability was assessed across the participant, predictor, and outcome domains. All 25 studies () were rated as having low concerns regarding applicability in these three domains according to PROBAST. Detailed assessments of risk of bias and applicability are presented in . 【Clinical impact and intervention effects of prediction models】 None of the included studies directly evaluated the effects of interventions guided by the developed prediction models. Although the studies reported model performance indicators such as AUC, sensitivity, specificity, calibration, and validation results, no study assessed whether applying the model in clinical practice improved enteral nutrition delivery or patient outcomes. Specifically, there was no evidence from the included studies that model-guided management reduced the incidence of enteral feeding intolerance, improved target energy achievement, reduced feeding interruption, aspiration, infection, ICU length of stay, or mortality. 【Discussion】 【Frequently reported predictors and heterogeneity across models】 The results of this review showed substantial heterogeneity in both candidate predictor selection and final predictor retention across the included models. No single predictor was consistently identified across all included studies. APACHE II score, blood glucose, mechanical ventilation, age, albumin, and intra-abdominal pressure were relatively frequently reported predictors, but each appeared only in a subset of models. Therefore, these predictors should be interpreted as recurrent signals or hypothesis-generating factors rather than definitive or universally applicable predictors of enteral feeding intolerance. Differences in study populations, EFI definitions, candidate predictor sets, variable selection methods, and validation strategies may all have contributed to the variability in predictors retained across models. Eleven studies (, , , , , ) identified APACHE II score as a predictor of enteral feeding intolerance. APACHE II is a global measure of physiological derangement and disease severity rather than a gastrointestinal-specific measure. Reviews of gastrointestinal dysmotility and f