Detecting Uncertainty of Mortality Prediction Using Confident Learning.

Area of Science:

• Medical Informatics
• Machine Learning in Healthcare
• Critical Care Medicine

Background:

• Early mortality prediction in intensive care units (ICUs) is crucial for patient management.
• Existing machine learning (ML) models and severity scores face challenges due to data uncertainty, including irregular sampling, missing data, and patient heterogeneity.
• Uncertainty quantification in ML predictions for critically ill patients remains an underexplored area.

Purpose of the Study:

• To incorporate sample-uncertainty information into ML-based mortality prediction models for ICU patients.
• To evaluate the performance and robustness of ML models augmented with confident learning (CL) for mortality prediction.
• To demonstrate the importance of uncertainty quantification in improving the reliability of patient outcome predictions.

Main Methods:

• Utilized confident learning (CL) to address sample uncertainty in mortality prediction.
• Applied CL to enhance state-of-the-art ML models.
• Evaluated model performance on a large dataset of 139,367 unique ICU admissions from the eICU Collaborative Research Database (eICU-CRD).

Main Results:

• Confident learning effectively incorporated sample-uncertainty information into mortality prediction models.
• ML models augmented with CL showed improved robustness against epistemic error.
• The study validated the significance of uncertainty quantification for enhancing patient outcome prediction accuracy in ICUs.
• CL-based models demonstrated better performance in handling class imbalance within the dataset.

Conclusions:

• Uncertainty quantification is vital for accurate and reliable early mortality prediction in ICUs.
• Confident learning offers a robust method to improve ML model performance by addressing data uncertainty and class imbalance.
• Augmenting ML models with CL leads to more dependable predictions for critically ill patients, aiding clinical decision-making.