ICU Open-Access Databases for Artificial Intelligence in Sepsis: Balancing Innovation With Data Quality Challenges.

Abstract

Objectives: To evaluate the suitability of three major open-access ICU databases (Medical Information Mart for Intensive Care IV [MIMIC-IV], eICU Collaborative Research Database [eICU-CRD], and Amsterdam University Medical Center Database [AmsterdamUMCdb]) for artificial intelligence (AI)-based sepsis research, with a focus on data completeness, internal consistency, terminological standardization, and structural redundancy affecting clinical interpretability and reproducibility.

Design: A descriptive comparative design was employed to evaluate data completeness, terminological consistency, structural integrity, and clinical usability, with emphasis on sepsis-related variables such as antimicrobial administration, hemodynamic support, and mortality indicators.

Setting: Open-access ICU electronic health record databases including MIMIC-IV (2008-2019), eICU-CRD (2014-2015), and AmsterdamUMCdb (2003-2016), accessed under credentialed use agreements. Documentation, schema, and metadata were reviewed in parallel to primary datasets.

Interventions: Structured queries and descriptive analyses were performed to quantify missing values, distinct diagnosis strings, medication administration completeness, and redundancy of mortality markers. Manual semantic inspection of terminology, units, and coding (International Classification of Diseases, 9th revision/10th revision) was applied to evaluate inconsistency and typographical variability. Comparisons were made across key clinical dimensions including diagnosis, mortality, microbiology, and antibiotic administration.

Measurements and main results: Marked heterogeneity and data quality limitations were identified. MIMIC-IV demonstrated extensive missingness in medication route (49%) and dose (58%), 99.9% missing fields in microbiology quantity, and diagnosis redundancy with 17,557 unique strings and 174 near-duplicate variants. eICU-CRD exhibited 1.9% admissions lacking diagnosis and discordance between ICU and hospital mortality in 7,319 stays. AmsterdamUMCdb showed fewer structural inconsistencies but limited demographics and the highest proportion of admissions without diagnosis (61.4%), complicated by mixed English-Dutch terminology. All three databases lacked reliable sepsis diagnosis timestamps, constraining analysis of early intervention.

Conclusions: Although widely used and valuable for sepsis research, open-access ICU databases exhibit substantial missingness, redundancy, semantic variability, and demographic imbalance. These limitations require rigorous preprocessing and highlight the need for standardized data collection to improve the validity and generalizability of AI-driven sepsis studies.