1,350 FDA Approvals Still Don't Prove What Our 55% Forecast Needs Them to Prove

Let's be precise about the forecast target, because precision is where this forecast lives or dies. 'Fully AI-driven diagnostic' means physician-free final decision authority — no mandatory human review before the result is acted upon clinically. IDx-DR cleared this bar for diabetic retinopathy in 2018. That precedent is real and we've accounted for it. Our forecast is therefore asking a narrower question: will the FDA extend autonomous diagnostic clearance beyond the single-specialty, single-condition precedent into a domain where the diagnostic stakes, liability exposure, and physician oversight norms are materially higher? Radiology for cancer screening. Cardiology. Pathology. That's the threshold that hasn't been crossed.

The 1,350 approval figure is proximate evidence at best. It proves the FDA has built administrative machinery for AI device review. It does not prove the agency has resolved the liability transfer question that actually blocks the next step. The dominant category in that 1,350 is radiology — AI tools that flag, triage, or augment — but where the radiologist still signs. Volume of approvals in the augmentation category tells us nothing about FDA willingness to approve in the autonomous category for high-stakes indications. These are philosophically and legally distinct questions, and conflating them is the evidence error we most need to avoid here.

What actually moves our probability is not the approval count — it's signals about FDA's framework evolution. The agency's action plans on AI/ML-based software as a medical device, and specifically any movement toward predetermined change control protocols that allow autonomous operation, matter more than raw clearance volume. We've seen the FDA signal openness to adaptive frameworks, which is why we moved from 52% to 55%. But 'signaling openness' is not 'establishing a pathway.' The AMA's position against removing physician oversight for final diagnostic decisions remains the real bottleneck — not model performance, not regulatory throughput.

Honestly, this is the part of our thesis that keeps us up at night: we may be too anchored to the regulatory pathway and underweighting the possibility that a company structures a product cleverly enough to achieve de facto autonomy within existing oversight frameworks — the physician 'review' becomes so nominal it's functionally meaningless. That would resolve our forecast YES without FDA ever formally endorsing the autonomous category. What would move us above 65%? A specific FDA guidance document addressing liability attribution for AI-primary diagnostics, or a cleared product where the mandated physician review window drops below 60 seconds for non-emergency indications. What would drop us below 40%? A major diagnostic AI error event that triggers Congressional hearings — one high-profile misdiagnosis could set the timeline back two years.