使用常规收集的电子健康记录数据预测院内谵妄的模型：系统综述。
完整摘要 · 中文译文
背景：谵妄是一种常见且具有重要临床意义的院内急性精神状态恶化形式。基于电子健康记录（EHR）的预测模型可能支持早期识别和针对性预防，但其方法学质量、验证严谨性和临床就绪程度仍不确定。 目的：本系统评价旨在综合并严格评估使用常规收集的EHR数据开发的院内谵妄预测模型，重点关注模型特征、验证策略、性能、偏倚风险和临床适用性。 方法：我们检索了PubMed、MEDLINE、Embase、PsycINFO和Web of Science，检索时间从建库至2025年11月11日。符合条件的研究开发、验证或评估了使用常规收集的EHR或管理数据预测成人住院期间急性精神状态恶化的多变量预测模型。尽管纳入标准较为宽泛，但所有纳入研究均将恶化操作化为谵妄。数据提取依据CHARMS（预测模型系统评价的严格评价和数据提取清单）和TRIPOD（个体预后或诊断多变量预测模型的透明报告）或TRIPOD-人工智能指南进行。模型性能、验证、校准和实施特征以叙述性方式综合。使用PROBAST（预测模型偏倚风险评估工具）评估偏倚风险和适用性。 结果：二十九项研究符合纳入标准。 这些证据可归类为4个相互重叠的预测任务：入院时或住院早期风险分层、围手术期或术后预测、重症监护病房动态预测，以及现有工具的外部验证或工作流程评估。大多数研究为回顾性队列研究（20/29，69%），均在普通病房、普通病房与重症监护病房混合病房、重症监护病房或急诊科环境中开展。机器学习或混合方法较为常见（18/29，62%），但更复杂的模型并未持续优于统计或基于规则的方法。在29项研究中，有24项（83%；受试者工作特征曲线下面积范围0.77-0.97）报告了内部区分度，而有12项研究报告了外部区分度，15项研究报告了校准度。仅有3项研究报告了决策曲线分析，前瞻性评估或工作流程整合仍然有限。总体偏倚风险为低风险的有8项研究，不清楚的有10项研究，高风险的有11项研究，主要原因是分析领域存在局限性。 结论：常规收集的电子健康记录数据可支持医院各环境下的谵妄风险预测，且许多模型显示出中等至较高的区分度。然而，尚无单一算法可随时用于常规临床应用。该领域仍受到预测任务异质性、结局判定不一致、校准度和决策分析报告薄弱，以及外部或前瞻性评估不足的限制。未来研究应在模型开发之前明确预期临床使用场景，评估校准度。 在部署前，应结合区分能力评估模型的校准度与临床实用性，并在不同机构、时间段及工作流程中对模型进行测试。
完整摘要 · 英文原文
BACKGROUND: Delirium is a common and clinically important form of acute in-hospital mental status deterioration. Electronic health record (EHR)-based prediction models may support early identification and targeted prevention, but their methodological quality, validation rigor, and clinical readiness remain uncertain. OBJECTIVE: This systematic review aimed to synthesize and critically evaluate prediction models for in-hospital delirium developed using routinely collected EHR data, focusing on model characteristics, validation strategies, performance, risk of bias, and clinical applicability. METHODS: We searched PubMed, MEDLINE, Embase, PsycINFO, and Web of Science from inception to November 11, 2025. Eligible studies developed, validated, or evaluated multivariable prediction models using routinely collected EHR or administrative data to predict acute mental status deterioration during adult hospital admissions. Although eligibility criteria were broad, all included studies operationalized deterioration as delirium. Data extraction was informed by CHARMS (Checklist for Critical Appraisal and Data Extraction for Systematic Reviews of Prediction Modeling Studies) and TRIPOD (Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis) or TRIPOD-artificial intelligence guidance. Model performance, validation, calibration, and implementation features were synthesized narratively. Risk of bias and applicability were assessed using PROBAST (Prediction Model Risk of Bias Assessment Tool). RESULTS: Twenty-nine studies met the inclusion criteria. The evidence clustered into 4 overlapping prediction tasks: admission or early-stay risk stratification, perioperative or postoperative prediction, dynamic intensive care unit prediction, and external validation or workflow evaluation of existing tools. Most studies were retrospective cohorts (20/29, 69%) and were conducted in general ward, mixed ward-intensive care unit, intensive care unit, or emergency department settings. Machine learning or hybrid approaches were common (18/29, 62%), but more complex models did not consistently outperform statistical or rule-based approaches. Of 29 studies, internal discrimination was reported in 24 (83%; area under the receiver operating characteristic curve range 0.77-0.97) studies, whereas external discrimination was reported in 12 studies and calibration in 15 studies. Decision curve analysis was reported in 3 studies, and prospective evaluation or workflow integration remained limited. Overall risk of bias was low in 8 studies, unclear in 10 studies, and high in 11 studies, mainly because of analysis-domain limitations. CONCLUSIONS: Routinely collected EHR data can support delirium risk prediction across hospital settings, and many models show moderate to high discrimination. However, no single algorithm is ready for routine adoption. The field remains limited by heterogeneous prediction tasks, inconsistent outcome ascertainment, weak calibration and decision-analytic reporting, and insufficient external or prospective evaluation. Future studies should define the intended clinical use case before model development, evaluate calibration and clinical usefulness alongside discrimination, and test models across institutions, time periods, and workflows before deployment.