人工智能与机器学习在分诊中的技术整合与传统分诊的比较及远程医疗在急诊和急症护理环境中用于患者优先排序的叙述性综合：系统综述与荟萃分析。
完整摘要 · 中文译文
背景：传统的分诊依赖于临床医生的判断和首次接触时可获得的有限信息。人工智能（AI）和机器学习（ML）可能改善早期风险分层，而远程分诊和面向患者的工具主要支持远程评估、可及性和就医路径引导。 目的：本研究旨在评估AI和ML辅助分诊系统的诊断和预后表现，并总结关于远程分诊、工作流程、实施、经济和公平性结局的证据。 方法：纳入2010年1月1日至2025年12月31日期间发表的、评估AI、ML、自然语言处理、远程医疗或数字化增强分诊的比较性研究。首次临床接触即纳入。检索了PubMed/MEDLINE、CINAHL、Web of Science、Scopus、预印本服务器、试验注册库和会议论文集。筛选和数据提取由两人独立重复进行。使用适合研究设计的工具评估偏倚风险。临床可比的受试者工作特征曲线下面积（AUC）估计值采用随机效应荟萃分析进行合成；异质性较大的诊断、操作、实施、经济和公平性结局进行叙述性总结。探索性情景建模预测了每10000例急诊科（ED）就诊的影响。 结果：共纳入27项研究，其中13项属于定量证据流，14项属于叙述性综合。4项评估危重症或病情恶化结局的成人ED研究显示，AI辅助分诊具有较高的合并判别能力。 isted triage (AUC 0.874; 95% CI 0.845-0.898; I²=94.4%; 95% prediction interval 0.803-0.922)。四项研究报告了可直接提取的研究内比较，均倾向于AI辅助模型（AUC差值分别为0.097、0.120、0.108和0.065），但由于无法获得配对方差估计，这些结果不能进行合并。入院/分诊处置研究报告的AI模型AUC范围为0.80至0.93，由于极端的异质性排除了有意义的合并，因此进行了描述性总结。两项儿科研究（AUC为0.85-0.99）进行了描述性总结。预测模型研究主要受限于模型分析、校准报告和外部验证方面的不足。叙述性证据表明，远程分诊改善了前端运营指标，包括报告的临床评估时间减少15%-40%，未就诊离开率降低2-8个百分点，而面向患者的症状检查器通常偏向敏感性而非特异性。探索性情景建模显示，在特定的患病率和阈值假设下，AI辅助分诊可能与每10,000次急诊就诊中漏诊危重疾病病例减少12-38例以及假阳性升级增加45-120例相关。经济和公平性发现是基于假设的预测，而非观察到的临床、成本效益或公平性结局。 结论：AI和ML辅助分诊模型在成人急诊风险分层中显示出良好的区分能力；然而，基于AUC的区分能力并不能确立在临床上的优效性。 具有临床意义的决策阈值、患者结局改善或临床实施准备度。解读受到显著异质性、不完整的阈值水平性能和校准报告、有限的外部验证以及稀缺的儿科证据的限制。在常规实施之前，需要进行前瞻性多中心验证、工作流程评估、亚组校准与公平性审计以及正式的经济学评价。
完整摘要 · 英文原文
BACKGROUND: Conventional triage relies on clinician judgment and limited information available at first contact. AI and machine learning (ML) may improve early risk stratification, whereas tele-triage and patient-facing tools primarily support remote assessment, access, and care routing. OBJECTIVE: This study aimed to evaluate the diagnostic and prognostic performance of AI- and ML-assisted triage systems and summarize evidence on tele-triage, workflow, implementation, economic, and equity outcomes. METHODS: Comparative studies published between January 1, 2010, and December 31, 2025, evaluating AI-, ML-, natural language processing-, telehealth-, or digitally enhanced triage. At first, clinical contacts were included. PubMed/MEDLINE, CINAHL, Web of Science, Scopus, preprint servers, trial registries, and conference proceedings were searched. Screening and data extraction were performed independently in duplicate. Risk of bias was assessed using design-appropriate tools. Clinically comparable area under the receiver operating characteristic curve (AUC) estimates were synthesized using random-effects meta-analysis; heterogeneous diagnostic, operational, implementation, economic, and equity outcomes were summarized narratively. Exploratory scenario modeling projected effects per 10,000 emergency department (ED) visits. RESULTS: Twenty-seven studies were included, comprising 13 studies in the quantitative evidence stream and 14 in the narrative synthesis. Four adult ED studies evaluating critical care or deterioration outcomes demonstrated high pooled discrimination for AI-assisted triage (AUC 0.874; 95% CI 0.845-0.898; I²=94.4%; 95% prediction interval 0.803-0.922). Four studies reported directly extractable within-study comparisons favoring AI-assisted models (AUC differences of 0.097, 0.120, 0.108, and 0.065), although these could not be pooled because paired variance estimates were unavailable. Admission/disposition studies reported AI-model AUCs ranging from 0.80 to 0.93 and were summarized descriptively because extreme heterogeneity precluded meaningful pooling. Two pediatric studies (AUCs of 0.85-0.99) were summarized descriptively. Prediction-model studies were primarily limited by deficiencies in model analysis, calibration reporting, and external validation. Narrative evidence suggested improved front-door operational metrics with tele-triage, including reported reductions of 15%-40% in time to clinical assessment and 2-8 percentage points in left-without-being-seen rates, whereas patient-facing symptom checkers generally favored sensitivity over specificity. Exploratory scenario modeling illustrated that, under specified prevalence and threshold assumptions, AI-assisted triage could be associated with 12-38 fewer missed critical illness cases and 45-120 additional false-positive escalations per 10,000 ED visits. Economic and equity findings were assumption-dependent projections rather than observed clinical, cost-effectiveness, or equity outcomes. CONCLUSIONS: AI- and ML-assisted triage models demonstrate promising discrimination for adult ED risk stratification; however, AUC-based discrimination does not establish superiority at clinically relevant decision thresholds, improved patient outcomes, or readiness for clinical implementation. Interpretation is limited by substantial heterogeneity, incomplete threshold-level performance and calibration reporting, limited external validation, and sparse pediatric evidence. Prospective multicenter validation, workflow assessment, subgroup calibration and equity auditing, and formal economic evaluation are required before routine implementation.