Synthetic Patient Data Generation and Evaluation in Disease Prediction Using Small and Imbalanced Datasets

Area of Science:

• Medical Informatics
• Artificial Intelligence
• Data Science

Background:

• Chronic non-communicable diseases require advanced management tools.
• Artificial Intelligence (AI) shows promise in medical diagnosis and prediction.
• Challenges in medical data include limited high-fidelity datasets and patient privacy concerns.

Purpose of the Study:

• To develop and evaluate a framework for synthetic medical data generation.
• To assess the feasibility of synthetic data for preserving data integrity.
• To determine the impact of synthetic data on Machine Learning classification performance.

Main Methods:

• A framework utilizing synthetic data generation algorithms was developed.
• Eight tabular medical datasets were used for testing.
• Statistical metrics evaluated synthetic data integrity; F1-score assessed classification performance.

Main Results:

• Synthetic data generation successfully preserved data integrity across various dataset sizes.
• Machine Learning classifiers trained on synthetic data achieved comparable performance to those trained on real data.
• Combining synthetic and real data further enhanced classification accuracy.

Conclusions:

• Synthetic data generation is a feasible solution for addressing medical data limitations.
• The proposed framework effectively generates synthetic data that maintains data integrity and classification utility.
• Synthetic data can augment real medical datasets, improving AI model development for disease management.