Improving the Performance of Outcome Prediction for Inpatients With Acute Myocardial Infarction Based on Embedding...

Area of Science:

• Medical informatics
• Machine learning in healthcare
• Clinical data science

Background:

• Electronic medical records (EMRs) are crucial for healthcare quality improvement.
• Representation learning from EMRs is gaining attention for extracting hidden health information.

Purpose of the Study:

• To propose an improved patient representation method.
• Enhance outcome prediction for acute myocardial infarction (AMI) inpatients.
• Incorporate feature associations and task-specific importance into patient representations.

Main Methods:

• Medical concepts (diagnoses, tests, medications) were embedded using an improved skip-gram algorithm with feature association strengths.
• Patients were represented by summing feature embeddings weighted by task-specific importance.
• Applied the representation to mortality risk prediction for AMI inpatients, comparing with reference methods using AUROC, AUPRC, and F1-score.

Main Results:

• The proposed embedding-based representation demonstrated superior predictive performance on both public and private datasets.
• Achieved higher mean AUROCs (0.878, 0.973), AUPRCs (0.220, 0.505), and F1-scores (0.376, 0.674) compared to reference methods.
• Integrated feature importance highlighted critical features for prediction and clinical practice.

Conclusions:

• Feature associations and importance enhance patient representation for improved prediction.
• The method contributes to better predictive model performance and interpretability.
• This approach effectively leverages EMR data for clinical outcome prediction.