AI Cancer Discovery, Safety Concerns, and FDA Changes - Dec 17, 2025

Today brings significant developments in AI healthcare. Six key stories shape our understanding of artificial intelligence's growing role in medicine:

Have suggestions? Reply to this email.

Harvard Medical School researchers found that AI pathology models learn to identify which hospital processed cancer slides, not just the cancer itself. The models showed remarkable accuracy in predicting the source hospital even when cancer information was removed from slides. This discovery raises concerns about AI reliability in clinical settings. Source

The finding suggests AI systems may make decisions based on irrelevant factors like slide preparation methods or scanning equipment rather than medical features. This could lead to biased diagnoses when AI models encounter samples from new hospitals or different processing protocols.

University of Sydney research warns that rapid AI integration in primary care clinics poses safety concerns. Digital scribes, ChatGPT, and other AI tools are entering GP practices without adequate testing or oversight. Source

The study highlights gaps between AI capabilities and real-world clinical needs. Researchers stress the need for proper validation before widespread deployment. Current AI tools may not handle complex medical scenarios that require human judgment and contextual understanding.

Mass General Brigham announced AIwithCare, a new company featuring software that uses generative AI to screen patients for clinical trial eligibility. The platform aims to streamline patient recruitment and improve trial participation rates. Source

Clinical trials often struggle to find eligible participants, delaying research and limiting access to experimental treatments. AI screening could identify suitable patients faster and more accurately than manual review processes.

The FDA received varied feedback on proposed AI performance monitoring requirements. Industry comments highlighted challenges with continual machine learning models that update automatically based on new data. Source

Companies expressed concerns about monitoring complexity and regulatory burden. Some argued for flexible approaches that account for different AI model types. Others supported strict oversight to ensure patient safety as AI systems evolve in clinical practice.

The NAACP released a blueprint addressing AI bias in healthcare, warning that artificial intelligence could worsen race-based health disparities. The organization calls for equity-focused AI development and deployment practices. Source

The blueprint emphasizes inclusive data collection, diverse development teams, and ongoing bias monitoring. It serves as a framework for healthcare organizations seeking to implement AI while protecting vulnerable populations from algorithmic discrimination.

University of Michigan researchers developed an AI model that helps diagnose cardiac amyloidosis, an under-recognized heart condition. The model, created with Invia Medical Imaging Solutions, analyzes medical imaging to identify this often-missed diagnosis. Source

Cardiac amyloidosis causes heart failure but frequently goes undiagnosed. Early detection enables targeted treatment and improves patient outcomes. The AI tool could help physicians catch this condition before irreversible heart damage occurs.

These developments highlight AI's rapid healthcare integration alongside growing awareness of implementation challenges. Success requires careful attention to bias, safety, and equity concerns while harnessing AI's diagnostic potential.