Interpretable Topic Features for Post-ICU Mortality Prediction.

Area of Science:

• Medical Informatics
• Computational Biology
• Public Health

Background:

• Electronic health records (EHRs) are vital for disease correlation and mortality studies.
• Post-discharge mortality analysis is crucial for timely intervention, cost reduction, and healthcare quality improvement.

Purpose of the Study:

• To develop a predictive model for post-discharge Intensive Care Unit (ICU) mortality.
• To enhance prediction by integrating clinical data with insights from medical notes using topic modeling.

Main Methods:

• Utilized electronic health records, incorporating International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) hierarchy.
• Employed Bayesian topic modeling, specifically Labeled-LDA, to extract topic features from clinical notes.
• Combined physiological measurements with extracted topic features for mortality prediction.

Main Results:

• Achieved high Area Under the Curve (AUC) values of 0.835 for 30-day and 0.829 for 6-month post-discharge mortality prediction.
• Demonstrated the effectiveness of topic features derived from Labeled-LDA in improving predictive accuracy.
• Highlighted the interpretability of topic features for understanding disease-mortality complexity.

Conclusions:

• The integration of topic features from EHRs significantly enhances post-discharge ICU mortality prediction.
• The developed model offers interpretable insights into the factors influencing mortality after hospital discharge.
• This approach holds potential for improving patient follow-up and healthcare management strategies.