AI news ·
Paragon Health Institute proposes voluntary framework to address AI medical device accuracy gaps
Paragon Health Institute proposed Digital Similarity Analysis, a framework that checks whether a patient's medical data matches what an AI device was trained on. When a mismatch is found, doctors get an alert to adjust how they use the tool.
New Framework Addresses AI Medical Device Safety Gap
Paragon Health Institute published a research paper proposing a framework to address a critical safety issue with AI-enabled medical devices: their unpredictable performance on real patients whose medical data differs from training data.
AI devices often work well during development testing but perform less reliably when deployed in clinical settings. This gap between test performance and real-world results is called generalization uncertainty. It's a central concern as AI for Healthcare becomes more prevalent in patient care.
The Cost Problem With Current Solutions
Existing approaches to manage this uncertainty rely on expert consultation and population-level risk assessments. These methods are expensive and create access disparities. Well-funded health systems can afford consultants; rural hospitals and safety-net providers cannot.
This divide threatens to concentrate AI benefits among wealthy institutions while leaving others behind.
Digital Similarity Analysis as an Alternative
Paragon proposes Digital Similarity Analysis (DSA), a voluntary framework that would evaluate whether an individual patient's medical information-X-rays, CT scans, mammograms, and similar data-matches what an AI device was trained on.
When DSA identifies a patient as an outlier, the physician receives an alert and can choose to: skip the device, request additional validation of its output, or use it with reduced confidence.
The approach shifts focus from broad demographic categories to individual patient characteristics. This shift may reduce algorithmic bias across different demographic groups.
Preserving Development Advantages
DSA allows device manufacturers to keep their training data confidential-a critical asset in Generative AI and LLM development. This differs from approaches that would require sharing proprietary information.
The framework doesn't eliminate generalization uncertainty entirely. Combined with postmarket surveillance, however, it could provide physicians with practical guidance while maintaining the technology's potential to improve patient care.
Paragon's work represents one of several recent policy papers on AI safety in healthcare, including studies on postmarket surveillance, regulatory guidelines, and cost reduction through AI applications.