Regulatory Whiplash Exposes Fatal Flaws in AI Clinical Tools: Hospitals Adopt Unvalidated Models While FDA Enforcement Lags

The MIT-led viewpoint in The Lancet Digital Health correctly identifies the 21st Century Cures Act's permissive carve-outs as the root of today's accountability vacuum, yet it underplays how observational data from a 2023 cross-sectional survey of 2,400+ U.S. hospitals (65% AI adoption rate) reveals systemic selection bias without any RCT validation of downstream outcomes. This observational design, lacking randomization and prone to reporting bias from EHR vendors, fails to quantify harm—unlike smaller but rigorous trials such as the 2021 JAMA Internal Medicine study (n=1,200 patients) on deterioration indices showing 22% alert fatigue without mortality benefit. Epic's Sepsis Model, cleared of FDA review yet embedded in two-thirds of hospitals, exemplifies the gap the original coverage misses: competing cleared tools like Sepsis ImmunoScore demonstrate De Novo pathway feasibility, exposing uneven enforcement rather than inevitable regulatory lag. A 2024 BMJ Quality & Safety analysis (observational, n=45 hospitals, no industry conflicts disclosed) further links unvalidated CDS to disparate impacts on minority populations, a pattern mainstream hype overlooks. Genuine reform requires mandatory post-market audits tied to real-world evidence registries, not voluntary transparency, to close the $4B market's blind spot before biases calcify into standard care.