🆕 AI triage noninferior: interval cancers 1.55 vs 1.76/1000

In a Swedish randomized, single-blinded, population-based non-inferiority trial (NCT04838756) of 105,915 women, AI-supported mammography screening was noninferior to standard double reading for interval cancers: 1.55 vs 1.76 per 1,000 (proportion ratio 0.88; 95% CI 0.65–1.18; p=0.41). Sensitivity was higher with AI (80.5% vs 73.8%; p=0.031) with identical specificity (98.5% in both), while reducing radiologist reading workload.

Why It Matters To Your Practice

• Interval cancers are a key safety metric in screening; this trial suggests AI triage can be deployed without increasing interval cancer rates (1.55 vs 1.76/1,000).

• Higher sensitivity with unchanged specificity implies potential for better detection without more false positives—important for patient anxiety, downstream testing, and clinic throughput.

• AI was used both to triage exams to single vs double reading and as detection support—closer to real-world workflow redesign than “AI as a second opinion” alone.

Clinical Implications

• Consider AI-supported triage models as an option to maintain safety while reducing double-reading burden, particularly where radiologist capacity is constrained.

• Expect sensitivity gains to be most evident for invasive cancers (reported as consistent across age and breast density), not for in-situ disease.

• Plan implementation with governance: define triage thresholds, escalation rules (single→double read), QA monitoring, and pathways for discordant reads.

Insights

• Noninferiority margin was 20%; observed proportion ratio favored AI (0.88) but with CI crossing 1.0—so interpret as “not worse,” not definitively better, for interval cancers.

• Interval cancer characteristics were descriptively less unfavorable with AI support (e.g., fewer invasive: 75 vs 89; fewer T2+: 38 vs 48; fewer non-luminal A: 43 vs 59), suggesting possible stage/biology shift.

• Specificity held steady at 98.5% in both arms (p=0.88), countering a common concern that AI-driven detection increases recalls.

The Bottom Line

• In this large Swedish Breast Cancer screening trial, AI-supported triage was noninferior for interval cancers (1.55 vs 1.76/1,000) with higher sensitivity and unchanged specificity—supporting real-world adoption where workload relief is needed.