Abstract
Regulating clinical artificial intelligence (AI) presents new challenges for policymakers worldwide. While existing rules may work for older AI technologies, new guidelines are needed for generative AI (GAI). This article looks at current U.S. strategies for regulating older clinical AI. It suggests a new approach to GAI regulation, stressing the need for quick action to enjoy AI’s benefits safely.
Introduction
AI has great potential in healthcare, making regulation necessary. Efforts are underway in the U.S., with President Biden’s executive order highlighting the need for AI controls. The European Union has also set up broad AI regulations. Despite extensive experience regulating healthcare technologies, AI’s complexity and wide use make it unique. This article focuses on AI used in direct patient care, with implications for healthcare globally.
Types of Clinical AI
Pregenerative Artificial Intelligence (PGAI)
PGAI includes several types of software, some of which have been regulated in the U.S. for decades. Key types include:
Clinical Decision Support (CDS): Shares evidence-based guidelines without analysing information.
Assisting Software: Helps interpret clinical data for decisions like insulin dosing or cardiovascular risk.
Machine Learning (ML): ML based PGAI learns from large datasets, finding patterns beyond human observation, like predicting diabetes risk from chest X-rays.
Generative Artificial Intelligence (GAI): GAI, unlike PGAI, is multipurpose and more like human intelligence. It uses large language models (LLMs) to process vast amounts of information and continually evolves. While promising, GAI’s ability to learn and improve raises concerns about reliability and transparency.
Regulation of PGAI
The U.S. has set up pathways to regulate ML-based PGAI, including:
Exempting CDS from Regulation: Because it only shares existing clinical data.
Software as a Medical Device (SaMD): FDA evaluates high-risk software for safety and effectiveness.
Predetermined Change Control Plan (PCCP): This plan allows ML software to improve through updates, balancing innovation with safety.
Challenges in Regulating GAI
GAI’s complexity and evolving nature make traditional regulations inadequate. Proposed solutions include:
Premarket Evaluation: Assess specific clinical uses of GAI using SaMD and PCCP pathways.
New Regulatory Approach: Treat GAI-based applications like clinical intelligence, requiring training, supervised use, regular updates, and public performance reporting.
Discussion
Regulating PGAI is possible with current frameworks, though more monitoring resources are needed. GAI, however, needs new regulatory methods, possibly treating it more like human intelligence. Effective GAI regulation is crucial to unlocking its benefits. Policymakers must refine regulatory science to ensure AI’s safe use in clinical practice, considering reimbursement and liability issues to guide responsible adoption.
In summary, the unique challenges posed by clinical AI require a proactive, adaptable regulatory framework. Embracing these challenges is essential for safely and effectively enhancing patient care with AI.
source: The Regulation of Clinical Artificial Intelligence | NEJM AI