Traumatic Brain Injury ICU Mortality Predictors from MIMIC-IV, 2008–2019

A retrospective analysis of 2,691 traumatic brain injury (TBI) patient records from the MIMIC-IV database between 2008 and 2019. The dataset, created by Ning Liu, contains clinical variables used to develop an interpretable machine learning model for predicting in-hospital mortality. The primary outcome occurred in 316 patients (11.7%), and a final set of 12 predictive variables was identified.

Use Cases

• Training mortality prediction models based on identified clinical predictors like Glasgow Coma Scale (GCS), age, and lab values.
• Analyzing non-linear risk relationships, such as the U-shaped association between body temperature and mortality risk mentioned in the study.
• Benchmarking machine learning algorithm performance (e.g., XGBoost, Random Forest) for clinical prediction tasks.
• Applying SHAP analysis for model interpretability to understand population-level and individual prediction drivers.

Strengths

• Includes 2,691 patient records with a clearly defined outcome (11.7% in-hospital mortality).
• Features were selected via a rigorous intersectional strategy (Univariate Logistic Regression, LASSO, Boruta).
• The final model (XGBoost) achieved a high AUROC of 0.873, as reported in the description.
• Data is licensed under CC-BY-4.0, permitting open sharing and adaptation.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for some modeling tasks.
• Data is sourced from a single database (MIMIC-IV) and time period, which may limit generalizability.

Provenance

Source

MIMIC-IV database

Collection Method

Retrospective analysis of clinical records

Time Range

2008–2019

Freshness

Last updated 2026-04-23 04:19:09; freshness should be verified.

Geography

null