Feature rearrangement based deep learning system for predicting heart failure mortality.

Area of Science:

• Medical Informatics
• Machine Learning
• Cardiology

Background:

• Heart failure (HF) mortality prediction is critical for patient care.
• Traditional models struggle with imbalanced HF data and poor feature representation.
• Accurate prediction aids in timely interventions and improved patient management.

Purpose of the Study:

• To propose a novel, fast, and accurate deep learning framework for HF mortality prediction.
• To address the challenges of data imbalance and feature representation in HF datasets.
• To enhance the clinical utility of predictive models for heart failure.

Main Methods:

• A feature rearrangement-based deep learning system was developed.
• A novel 'Feature rearrangement based convolutional layer' was introduced.
• The framework incorporates techniques to handle data imbalance, such as Focal loss.

Main Results:

• The proposed framework achieved superior performance in HF mortality prediction compared to existing methods.
• Highest average accuracy and area under the curve were observed for in-hospital, 30-day, and 1-year mortality.
• The study identified top 12 essential clinical features for HF prediction.

Conclusions:

• The developed deep learning method accurately predicts HF mortality across different timeframes.
• Feature rearrangement and Focal loss significantly improved prediction accuracy, especially in imbalanced datasets.
• The framework offers a robust pipeline for modeling electronic health records (EHRs) and managing imbalanced medical data.