Event Capture Rates for In-Hospital Mortality in ICU AKI Patients

A dataset of 81,876 ICU admissions from MIMIC-III/IV and 140,237 from eICU used to evaluate a deep-learning anomaly signal for predicting kidney replacement therapy and mortality. The data includes 381,700 time-series instances internally and 494,684 externally, built from seven-step daily creatinine-eGFR series. It was created by Yoonjin Kang and last updated in May 2026.

Use Cases

• Predicting 24-96 hour kidney replacement therapy based on creatinine-eGFR trajectory anomaly scores.
• Stratifying mortality risk in ICU patients using anomaly detection combined with KDIGO staging.
• Comparing the performance of anomaly-based detection against conventional KDIGO criteria for event capture.
• Validating time-series deep learning models for clinical outcome prediction across different ICU databases.

Strengths

• Includes data from two large, established ICU databases: MIMIC-III/IV (81,876 admissions) and eICU (140,237 admissions).
• Contains a substantial volume of time-series instances: 381,700 internally and 494,684 externally.
• Provides evaluation metrics (AUROC, accuracy, F1-score) for KRT and mortality prediction at multiple time horizons.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count for the primary dataset file is unknown, which may limit suitability assessment.
• Data may reflect the specific biases inherent to the MIMIC and eICU source databases.

Provenance

Source

Medical Information Mart for Intensive Care (MIMIC-III/IV) and the eICU Collaborative Research Database.

Collection Method

Seven-step daily creatinine–eGFR time series instances were built from patient records.

Freshness

Last updated 2026-05-13 17:34:46; freshness should be verified.