Heart Failure Diagnosis and Severity Estimation enhanced by Generative Adversarial Network.

Area of Science:

• Cardiology
• Machine Learning
• Artificial Intelligence

Background:

• Heart Failure (HF) management requires accurate patient classification, often challenged by subjective evaluations.
• Existing datasets may be insufficient for robust machine learning model training.
• The Kardiatool project provides comprehensive data from 487 subjects.

Purpose of the Study:

• To enhance classification performance for Heart Failure (HF) diagnosis and severity estimation.
• To investigate the utility of synthetic data generated by Generative Adversarial Models (GAMs).
• To improve patient management through automated classification, specifically NYHA class.

Main Methods:

• Generated synthetic data using a class-conditioned Generative Adversarial Model.
• Augmented real patient data with generated data.
• Employed eight conventional machine learning classifiers within a training-test framework.
• Utilized a dataset comprising demographic, laboratory, medication, risk factor, medical history, and physiological information.

Main Results:

• Achieved 95.97% accuracy for HF diagnosis and 90.23% for HF severity estimation using generated data.
• Observed an average performance increase across classifiers by 2.43% (accuracy) and 2.47% (F1-score) for diagnosis.
• Demonstrated an average performance increase across classifiers by 7.39% (accuracy) and 8.78% (F1-score) for severity estimation.

Conclusions:

• Synthetic data generation via GAMs significantly enhances machine learning model performance for HF classification tasks.
• Automated patient classification using enhanced datasets can aid in objective HF patient management.
• Interpretability of models trained with synthetic data is crucial for clinical adoption.