Feature Ranking in Predictive Models for Hospital-Acquired Acute Kidney Injury

Area of Science:

• Nephrology
• Medical Informatics
• Data Science

Background:

• Acute Kidney Injury (AKI) is a frequent and costly complication in hospitalized patients.
• Early AKI prediction is crucial due to the lack of specific treatments post-onset.
• Feature selection (FS) is vital for developing accurate AKI prediction models, but its robustness for AKI has not been studied.

Purpose of the Study:

• To compare the robustness and applicability of eight common FS methods for AKI prediction.
• To examine the heterogeneity in feature rankings generated by different FS methods.
• To assess the impact of FS method choice on AKI prediction performance and model reproducibility.

Main Methods:

• Utilized nine years of electronic medical records (EMR) data.
• Applied and compared eight distinct feature selection methods.
• Evaluated FS methods based on stability, selection result similarity, and prediction performance.

Main Results:

• Prediction accuracy was largely consistent across different FS methods.
• Feature importance rankings varied significantly among the evaluated FS methods.
• A positive correlation was found between FS method complexity and the required sample size.

Conclusions:

• Feature selection stability is critical for reproducible AKI prediction models.
• Identifying stable, important AKI risk factors is essential for clinical investigation.
• Robust FS methods can facilitate earlier and more reliable prediction of AKI.