An Interpretable Intensive Care Unit Mortality Risk Calculator.

Area of Science:

• Medical Informatics
• Clinical Decision Support
• Machine Learning in Healthcare

Background:

• High mortality risk for patients post-intensive care unit (ICU) discharge necessitates accurate prediction models.
• Current machine learning models offer high accuracy but lack interpretability for clinicians.
• Clinicians need insights into patient health and factors influencing mortality risk.

Purpose of the Study:

• To develop and compare machine learning models for predicting ICU patient mortality risk.
• To identify and benchmark the most salient predictive features across different machine learning techniques.
• To enhance the interpretability of mortality risk predictions for clinical decision-making.

Main Methods:

• Utilized patient profiles from the MIMIC-III clinical database.
• Constructed risk calculators using logistic regression, decision trees, random forests, k-nearest neighbors, and multilayer perceptrons.
• Conducted an extensive benchmarking study to compare feature importance across models.

Main Results:

• Observed a high degree of agreement in salient features across various machine learning methods.
• Consistently identified age, blood urea nitrogen level, and cardiac surgery recovery unit discharge as key predictors of mortality risk.
• Demonstrated the potential for machine learning to provide interpretable insights into patient mortality.

Conclusions:

• Machine learning models can effectively predict ICU patient mortality risk.
• Key demographic and clinical factors like age and blood urea nitrogen are crucial for risk assessment.
• This research facilitates better clinical interpretation of mortality risk predictions, supporting improved patient care.