Improving Explainability in Clinical Mortality Prediction Using Stacking Classifiers over Annotated Clinical Notes.

Area of Science:

• Clinical informatics
• Artificial intelligence in medicine
• Health services research

Background:

• Heart failure is a major cause of hospitalization and mortality.
• Early identification of high-risk patients is crucial for effective care.
• Predictive models for heart failure outcomes are needed.

Purpose of the Study:

• To develop and evaluate an interpretable machine learning model for predicting 3-month mortality after hospital discharge in heart failure patients.
• To assess the performance of a stacking classifier utilizing information from automatically annotated clinical notes.
• To compare the interpretability and accuracy of the stacking classifier against a global text classifier.

Main Methods:

• A stacking classifier was developed, with sub-models focusing on specific clinical entity types.
• A meta-classifier aggregated outputs from sub-models for a final mortality prediction.
• Real-world clinical notes were used for model training and evaluation.
• Performance was measured using Receiver Operating Characteristic Area Under the Curve (ROC-AUC).

Main Results:

• The stacking classifier achieved a ROC-AUC of 0.76.
• This performance was competitive with a global text classifier (ROC-AUC=0.78).
• The proposed method demonstrated improved interpretability compared to the global classifier.

Conclusions:

• Stacking classifiers offer a promising approach for interpretable prediction of heart failure mortality.
• Leveraging annotated clinical notes can enhance the clinical utility of predictive models.
• This methodology can aid clinicians in identifying high-risk patients for improved care and outcomes.