Scholarlink

Introduction <p>Electronic health records (EHRs) in critical care settings generate vast amounts of data that increasingly drive machine learning (ML) models for clinical decision support, yet data quality issues may have profound consequences for downstream prediction, classification, and optimization applications. This study aims to systematically examine EHR data quality issues in critical care medicine and their impact on ML model performance, clinical outcomes, and patient safety.</p> Methods <p>We conducted a systematized review following expert-based questions, searching MEDLINE, Embase, IEEE Xplore, ACM Digital Library, CINAHL, Google Scholar, DBLP, Web of Science, and the Cochrane Library. Six distinct questions addressed missing data patterns, temporal data quality, bias and health equity, multi-modal integration, real-time monitoring, and institutional variability.&#xa0;</p> Results <p> 281 relevant studies examining EHR data quality in critical care settings. After applying the eligibility criteria, 29 studies were selected. EHR data quality issues in critical care were pervasive and multifaceted. Missing data rates exceeded 80% for some variables, with 40% of predictive features being missingness indicators rather than actual values. EHR-related medication errors comprised 34% of all medication errors in ICUs, with one-third having life-threatening potential. Copy-paste prevalence reached 82% in residents’ progress notes. ML model performance degraded significantly under real-world conditions, with external validation showing AUC drops from 0.76 to 0.63 for sepsis detection models. Temporal data quality deteriorated throughout ICU stays, with vital sign quality degrading at 60–75% of average length of stay.</p> Conclusion <p>Data quality issues in critical care EHRs create cascading effects that compromise ML model reliability, clinical decision-making, and patient safety. The evidence demonstrates an urgent need for systematic data quality monitoring, bias-aware assessment methods, and comprehensive quality improvement frameworks specifically designed for critical care environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Discovery of data quality issues in electronic health records: profound consequences for critical care medicine applications – a systematized review

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年1月8日

查看全文
  • João Brainer Clares de Andrade,
  • Marconny Alexandre Oliveira de Medeiros Cavalcante,
  • Thiago Luís Marques Lopes,
  • João Marcos Secundino Treigher,
  • Mateus Dutra Balsells,
  • Júlia Lima Vasconcelos,
  • Lis Cavalcante Monteiro,
  • Déborah Danna da Silveira Mota

摘要

Introduction

Electronic health records (EHRs) in critical care settings generate vast amounts of data that increasingly drive machine learning (ML) models for clinical decision support, yet data quality issues may have profound consequences for downstream prediction, classification, and optimization applications. This study aims to systematically examine EHR data quality issues in critical care medicine and their impact on ML model performance, clinical outcomes, and patient safety.

Methods

We conducted a systematized review following expert-based questions, searching MEDLINE, Embase, IEEE Xplore, ACM Digital Library, CINAHL, Google Scholar, DBLP, Web of Science, and the Cochrane Library. Six distinct questions addressed missing data patterns, temporal data quality, bias and health equity, multi-modal integration, real-time monitoring, and institutional variability. 

Results

281 relevant studies examining EHR data quality in critical care settings. After applying the eligibility criteria, 29 studies were selected. EHR data quality issues in critical care were pervasive and multifaceted. Missing data rates exceeded 80% for some variables, with 40% of predictive features being missingness indicators rather than actual values. EHR-related medication errors comprised 34% of all medication errors in ICUs, with one-third having life-threatening potential. Copy-paste prevalence reached 82% in residents’ progress notes. ML model performance degraded significantly under real-world conditions, with external validation showing AUC drops from 0.76 to 0.63 for sepsis detection models. Temporal data quality deteriorated throughout ICU stays, with vital sign quality degrading at 60–75% of average length of stay.

Conclusion

Data quality issues in critical care EHRs create cascading effects that compromise ML model reliability, clinical decision-making, and patient safety. The evidence demonstrates an urgent need for systematic data quality monitoring, bias-aware assessment methods, and comprehensive quality improvement frameworks specifically designed for critical care environments.