Synthetic data, synthetic trust: navigating data challenges in the digital revolution.

Area of Science:

• Artificial Intelligence (AI)
• Medical AI
• Health Informatics

Background:

• The increasing reliance on artificial intelligence (AI) in healthcare is often fueled by the assumption that larger datasets improve model performance.
• Synthetic data generation is widely used to augment real-world data, addressing shortages but posing risks of bias and reduced generalizability.
• The rapid adoption of synthetic data in medical AI has led to 'synthetic trust,' an unfounded confidence in models that may lack clinical validity or demographic representation.

Purpose of the Study:

• To advocate for a cautious approach to using synthetic data in training clinical AI algorithms.
• To propose actionable safeguards for the responsible development and deployment of synthetic medical AI.
• To ensure end-to-end accountability, data integrity, and fairness in healthcare AI applications utilizing synthetic data.

Main Methods:

• Review of current practices in synthetic data augmentation for medical AI.
• Identification of potential risks associated with the overuse of synthetic data, including bias propagation and model degradation.
• Development of a framework for safeguards, encompassing data standards, testing protocols, and deployment disclosures.

Main Results:

• Overuse of synthetic data can lead to propagated biases, accelerated model degradation, and compromised generalizability.
• The phenomenon of 'synthetic trust' highlights the danger of unwarranted confidence in AI models trained on synthetic data.
• Proposed safeguards include establishing clear standards for training data, implementing fragility testing, and mandating disclosures of synthetic data origins.

Conclusions:

• Caution is essential when employing synthetic data for clinical AI model training.
• Implementing proposed safeguards can mitigate risks associated with synthetic data, ensuring data integrity and fairness.
• These measures establish new standards for the responsible and equitable use of synthetic data in healthcare applications.