Cancers detected at screening
Per 1,000 women · 95% confidence intervals
Loading chartFrom model scores
to clinical evidence
Screening outcomes, workload and physician decisions
More cancers found, fewer readings
Loading chart44.2% fewer readings
Reported total readings · slightly different arm sizes
What changed inside the screening pathway
The detected cancers
per 1,000 screened women · invasive and in situLoading chartFollow-up examinations
per 1,000 screened women · overlapping measuresLoading chartFalse positives are part of recalls, not additional cases. Reported differences in recall and false-positive rates were not statistically significant.
AI-supported: 53,043 women · 1,110 recalls · 772 false positivesStandard double reading: 52,872 women · 1,027 recalls · 765 false positives
What happened between screening rounds?
Interval cancers
Per 1,000 women · 95% CI · lower is betterLoading chartNon-inferiority established; a reduction was not demonstrated.
Sensitivity and specificity
Reported percentages · 95% CI in tooltipsLoading chartSensitivity: cancer cases detected. Specificity: non-cancer cases correctly excluded.
UNITED STATES · 50 PHYSICIANS · HISTORICAL GPT-4 TRIAL
A capable model is only part of the workflow
Median reasoning score
% of diagnostic-reasoning rubric pointsLoading chartThe adjusted difference
pp · estimate and 95% confidence intervalLoading chartThe interval crosses zero; a physician benefit was not established.
