AI Model Trained on 5.24 Million Routine Clinical Scans Outperforms Traditional Approaches, Raising Regulatory Questions

A July 2026 study in Nature Medicine reveals that a neuroimaging AI model trained on 5.24 million routine clinical scans outperforms traditional models trained on curated trial datasets. While messy, real-world data helps models generalize across diverse patient populations, it exposes a major regulatory gap. The FDA's current frameworks, including the December 2025 Real-World Evidence update and Predetermined Change Control Plans, lack clear validation pathways for routine-data-trained AI, leaving sponsors and CROs navigating an uncertain clearance landscape.

Routine clinical data AI superiority

• ▪The neuroimaging visual foundation model trained on routine clinical data demonstrated state-of-the-art diagnosis and enabled preliminary report generation and triage in real health systems.
• ▪A study published on July 31, 2026, demonstrated that a neuroimaging AI model trained on 5.24 million routine clinical CT and MRI scans outperformed models trained on traditional curated datasets.

Real-world data generalization

• ▪Research published in PLOS Medicine showed that dermatology AI algorithms trained on non-representative datasets had reduced diagnostic accuracy on darker skin tones compared to lighter skin tones.
• ▪Neuroimaging AI models trained on diverse, routine clinical data generalize better because they learn from real-world variations in scanner types, technician protocols, and patient demographics.

FDA regulatory framework gaps

• ▪The FDA Predetermined Change Control Plan rule allows AI developers to pre-specify post-clearance model updates, but it does not clarify how to validate routine-data-trained models for initial clearance.
• ▪The FDA updated its guidance on Real-World Evidence for medical devices in December 2025, expanding how sponsors may use Real-World Data to support regulatory decisions without always requiring randomized controlled trials.
• ▪The joint FDA, Health Canada, and MHRA guiding principles on Good AI Practice in Drug Development emphasize transparency in data sourcing and representativeness, but do not establish concrete submission requirements.

Clinical trial infrastructure implications

• ▪Sponsors who invest heavily in controlled, homogeneous data acquisition protocols may face competitive liabilities because their models are too brittle to perform under real-world clinical conditions.
• ▪The emerging clinical research model shifts the role of clinical trials from primary data generation sources to validation instruments for models trained on health system data partnerships.

Health system AI deployment trends

• ▪An ONC data brief found that 71% of hospitals reported using predictive AI integrated with their EHR systems in 2023-2024, up from 66% the prior year.
• ▪The global AI in medical imaging market was valued at USD 3.39 billion in 2024, with neurology holding the largest application segment.

2 sources