Can a normal mammogram predict breast cancer three years ahead? A UK model separated risk—but is not ready to set screening intervals
BREST-AI outperformed the Mirai comparator in a 7,596-woman UK case-control validation, but its predictive values were prevalence-adjusted estimates. The model still needs prospective workflow, calibration, equity and workload testing before clinical use.
The short version
- What happened
- The model was developed using 65,419 women and externally compared with Mirai in 1,899 UK cancer cases matched to 5,697 controls whose index mammograms had been read as normal.
- Who it affects
- Risk flags could eventually help target extra imaging or shorter follow-up, but false positives may also create recalls, biopsies and anxiety; this study did not measure that trade-off in practice.
- What remains unproven
- Case-control validation rather than prospective screening
Keep in mind
Case-control validation rather than prospective screening
Predictive values projected to population prevalence
Raw probabilities required recalibration
Small, underpowered US comparison
No UK ethnicity data or adequately powered fairness assessment
No patient outcomes, workload or cost-effectiveness measurement
Source analysis · 1 linked source. Primary analysed source date: 27 October 2026.
Read the full report · 4 min The evidence, methods, implications and limitations
The model separated future cases better in the primary UK test
A peer-reviewed study reports that BREST-AI can extract a three-year risk signal from screening mammograms that were read as normal. The model-development programme used 65,419 women; 58,591 were used for training, including 9,020 cancer cases, 1,633 benign cases and 47,938 women who remained cancer-free for three years. Its primary external UK validation used 1,899 later cancer cases and 5,697 age-, scanner- and site-matched controls from two separate NHS screening sites.
BREST-AI achieved an area under the receiver-operating curve of 0.751 (95% confidence interval 0.738–0.764), compared with 0.700 (0.687–0.714) for Mirai, an existing mammography risk model. AUC measures ranking across cases and controls; it does not say that 75.1% of individual predictions were correct or show that changing a woman's screening schedule improves health outcomes.[1]
Cross-vendor performance weakened, and the US comparison was uncertain
On a separate UK Siemens cohort of 400 cases and 1,200 matched controls, BREST-AI's AUC fell to 0.719 but remained above Mirai's 0.685. That drop matters because scanner and site shifts are a common route by which imaging models lose performance. In a much smaller US test of 97 cases and 291 controls, BREST-AI reached 0.670 versus 0.633 for Mirai; the difference was not statistically significant, and the authors call the comparison exploratory and underpowered.
The UK datasets did not include race or ethnicity, and the US subgroup confidence intervals overlapped. The paper therefore does not establish equitable performance across population groups. Broader validation would need enough participants to estimate calibration, sensitivity and predictive value by age, density, ethnicity, vendor and site—not merely an overall AUC.[1]
The distinction that matters
Three external tests, three different questions
The model was compared with Mirai using case-control data from different sites and equipment.
UK primary
1,899 cases + 5,697 controls
AUC 0.751 vs 0.700
UK Siemens
400 cases + 1,200 controls
AUC 0.719 vs 0.685
US exploratory
97 cases + 291 controls
AUC 0.670 vs 0.633; difference uncertain
The 8% predictive value is a projection, not a screening result
Because the external tests deliberately matched one case to three controls, they did not reproduce the prevalence of cancer in the English screening programme. The researchers reweighted results to eight cancers per 1,000 examinations. At a threshold set to 99% specificity, they estimated an 8.0% positive predictive value and a relative risk of 10.1 for the flagged group. Raw probabilities were substantially miscalibrated without adjustment, so local recalibration would be necessary before treating scores as absolute risk.
The model has not prospectively assigned extra imaging or shorter intervals, and the study measured no recalls, biopsies, anxiety, cost, interval cancers prevented or mortality benefit. The authors declared no competing interests and no specific funding. The evidence justifies a prospective NHS study with workload and equity safeguards; it does not justify changing a patient's screening interval today.
At the projected 8.0% positive predictive value, most people above the chosen high-risk threshold would still not develop cancer within three years. That does not make the signal useless: screening tools often trade sensitivity, specificity and workload. It does mean that any programme would need to specify what follows a flag and compare earlier detection with extra imaging, false alarms and capacity displaced from other patients.[1]