A deep learning system for heart failure mortality prediction

Area of Science:

• Cardiology
• Artificial Intelligence
• Medical Informatics

Background:

• Heart failure (HF) represents the advanced stage of numerous cardiac conditions.
• Prognosis for HF patients shows significant variability in mortality rates (5% to 75%).
• Accurate all-cause mortality evaluation is crucial for improving HF patient outcomes.

Purpose of the Study:

• To develop a deep learning system for accurate heart failure mortality prediction.
• To overcome limitations of traditional machine learning models with missing, high-dimensional, and imbalanced HF data.

Main Methods:

• An indicator vector approach was proposed to handle missing values and expand data dimensions.
• A convolutional neural network with varied kernel sizes was employed for feature extraction.
• A multi-head self-attention mechanism was utilized for comprehensive channel information capture.
• The focal loss function was implemented to effectively manage data imbalance.

Main Results:

• The system demonstrated effective and rapid prediction of four distinct mortality timeframes: within 30, 180, 365 days, and after 365 days.
• Data from the MIMIC-III database, including 10,311 patients, was utilized for system validation.
• Deep SHAP interpretation identified the top 15 predictive characteristics, confirming the system's efficacy.

Conclusions:

• The proposed deep learning system offers a robust solution for predicting heart failure mortality.
• The system's ability to handle data complexities and its interpretability enhance its clinical utility.
• Findings support the system's potential to improve medical services and patient management in cardiology.