人工智能辅助临床决策支持在初级保健中的真实世界有效性：系统综述。
完整摘要 · 中文译文
背景：人工智能驱动的临床决策支持系统（CDSS）正被用于初级保健，但其对临床决策和患者结局的影响仍不确定。 目的：本研究旨在综合临床医生面向人工智能驱动的CDSS在初级保健中的真实世界评估，并评估其在诊断性能、诊疗改变、临床医生效率及患者重要结局方面的证据。 方法：遵循PRISMA（系统评价和荟萃分析首选报告项目）2020指南，检索了5个主要文献数据库（PubMed、Scopus、Web of Science核心合集、Cochrane CENTRAL和CINAHL Ultimate）、一个补充的EBSCOhost平台搜索以及2个试验注册库，检索时间为2011年至2025年。纳入的研究评估了由临床医生在常规初级或门诊医疗中使用的学习型（数据衍生）或基于病例的人工智能组件，并报告了临床、行为、过程、效率或诊断准确性结局。筛选由2名评审员独立进行。偏倚风险使用适合设计的方法评估，并对4个结局体系应用了GRADE（推荐分级评估、制定与评价）。确定性基于规则的工具仅作为情境对照保留。 结果：共对66篇报告进行了全文评估，24项研究符合综述标准；其中10项为人工智能驱动并构成评估的综述人群，而14项确定性研究仅作为情境对照。检测结果不一致：人工智能心电图增加了新发低射血分数诊断 比值比（OR）为1.32，95%置信区间（CI）1.01-1.61，而单独使用的电子健康记录机器学习痴呆标志物则未显示此关联（校正后OR 0.84，95% CI 0.63-1.11）。诊断性研究显示黑色素瘤辨别力较高（受试者工作特征曲线下面积[AUROC] 0.960），但青光眼辨别力仅为中等（AUROC 0.80；敏感性65%）。非随机前后对照研究报告了尿路感染治疗成功率的提高（从75%增至80%），并改变了4名全科医生中3名的糖尿病视网膜病变筛查标准。一项暴露前预防试验总体结果为阴性；一个多组分跌倒预防临床决策支持系统（CDSS）改善了共享决策，但药物调整效果尚无定论。临床医生报告在哮喘护理中预估病历查阅时间有所减少。两项评估患者重要结局的试验均未显示人工智能临床决策支持系统（AI-CDSS）的获益；一项肌肉骨骼功能结局更倾向于常规护理。 结论：截至2025年12月发表的证据显示，初级保健中人工智能驱动的CDSS效果不一。一些研究报告了检测或护理流程的改善，但未证实对患者重要的获益。异质性和有限的结果测量阻碍了对这些差异的因果解释。未来试验应同时评估临床决策和患者结局。 试验注册：PROSPERO CRD420261302104；https://www.crd.york.ac.uk/PROSPERO/view/CRD420261302104。
完整摘要 · 英文原文
BACKGROUND: AI-enabled clinical decision support systems (CDSS) are being used in primary care, but their effects on clinical decisions and patient outcomes remain uncertain. OBJECTIVE: This study aimed to synthesize real-world evaluations of clinician-facing AI-enabled CDSS in primary care and assess evidence for diagnostic performance, changes in care, clinician efficiency, and patient-important outcomes. METHODS: Following PRISMA (Preferred Reporting Items for Systematic Reviews and Meta-Analyses) 2020, 5 principal bibliographic databases (PubMed, Scopus, Web of Science Core Collection, Cochrane CENTRAL, and CINAHL Ultimate), a supplementary EBSCOhost platform search, and 2 trial registries were searched for publications from 2011 through 2025. Eligible studies evaluated a learned (data-derived) or case-based AI component used by clinicians in routine primary or ambulatory care and reported a clinical, behavioral, process, efficiency, or diagnostic accuracy outcome. Screening was performed independently by 2 reviewers. Risk of bias was assessed with design-appropriate tools, and GRADE (Grading of Recommendations Assessment, Development and Evaluation) was applied to 4 outcome bodies. Deterministic rule-based tools were retained only as contextual comparison. RESULTS: A total of 66 reports were assessed at full text, and 24 studies met the review criteria; 10 were AI-enabled and formed the appraised review population, while 14 deterministic studies were contextual only. Detection findings were inconsistent: AI electrocardiography increased new low ejection fraction diagnoses (odds ratio [OR] 1.32, 95% CI 1.01-1.61), whereas an electronic health record machine learning dementia marker used alone did not (adjusted OR 0.84, 95% CI 0.63-1.11). Diagnostic studies showed high melanoma discrimination (area under the receiver operating characteristic curve [AUROC] 0.960) but only moderate glaucoma discrimination (AUROC 0.80; sensitivity 65%). Nonrandomized before-after studies reported increased urinary tract infection treatment success (from 75% to 80%) and changed diabetic retinopathy screening criteria in 3 of 4 general practitioners. A pre-exposure prophylaxis trial was null overall; a multicomponent fall prevention CDSS improved shared decision-making, with inconclusive medication-change effects. Clinicians reported lower estimated record review time in asthma care. Neither trial assessing patient-important outcomes demonstrated benefit from AI-CDSS; one musculoskeletal functional outcome favored usual care. CONCLUSIONS: Evidence published through December 2025 showed mixed effects of AI-enabled CDSS in primary care. Some studies reported improvements in detection or care processes, but patient-important benefit was not demonstrated. Heterogeneity and limited outcome measurement prevented causal explanations for these differences. Future trials should assess clinical decisions and patient outcomes together. TRIAL REGISTRATION: PROSPERO CRD420261302104; https://www.crd.york.ac.uk/PROSPERO/view/CRD420261302104.