A New Insight Into Missing Data in Intensive Care Unit Patient Profiles: Observational Study

Area of Science:

• Medical Informatics
• Clinical Data Science
• Predictive Analytics

Background:

• Substantial and unavoidable data gaps exist in intensive care unit (ICU) patient profiles.
• Data incompleteness in electronic health records (EHRs) may not always be random.

Purpose of the Study:

• Investigate hidden information within missing data in ICU EHRs.
• Determine if data absence itself predicts patient health status.
• Assess the predictive power of missingness indicators for mortality.

Main Methods:

• Utilized the Medical Information Mart for Intensive Care III (MIMIC-III) database.
• Introduced missingness indicators for laboratory tests (LTs).
• Applied filter and embedded feature selection methods.
• Evaluated prediction models (logistic regression, decision tree, random forest) using observed data and missingness indicators.

Main Results:

• Missingness indicators constituted over 40% of selected predictors across mortality types and ICU days.
• Missingness indicators alone achieved reasonable mortality prediction (e.g., AUROC of 0.6836±0.012 for 30-day mortality).
• Incorporating missingness indicators improved prediction model performance, with a maximum AUROC increase of 0.0426.

Conclusions:

• The presence or absence of LT measurements is informative and predicts in-hospital and 30-day mortality.
• Missing data indicators offer statistically significant prediction improvement.
• Missing data may reflect clinician judgment, providing predictive power beyond observed values.