Table 2.
AI performance on the internal test set across different sub-populations. We reported the AUROC of the AI system with 95% confidence intervals on the internal test set. The biopsied population only includes exams where at least one biopsy was recommended. We stratified exams based on patient age, mammographic breast density, and the manufacturer of the US devices. Mammographic breast density was categorized based on the BI-RADS standards69.
| Population | AUROC (95% CI) | No. of breasts | No. of cancers |
|---|---|---|---|
| Overall population | 0.976 (0.972, 0.980) | 79,078 | 1248 |
| Biopsied population | 0.940 (0.934, 0.947) | 12,973 | 1248 |
| Age | |||
| < 40 yrs old | 0.969 (0.955, 0.982) | 5176 | 72 |
| 40 − 49 yrs old | 0.970 (0.955, 0.986) | 19,677 | 160 |
| 50 − 59 yrs old | 0.981 (0.975, 0.986) | 24,142 | 292 |
| 60 − 69 yrs old | 0.980 (0.973, 0.985) | 19,039 | 326 |
| ≥70 yrs old | 0.969 (0.958, 0.981) | 11,044 | 398 |
| Breast density | |||
| Entirely fatty | 0.964 (0.942, 0.983) | 1157 | 54 |
| Scattered fibroglandular densities | 0.975 (0.961, 0.982) | 19,199 | 441 |
| Heterogeneously dense | 0.979 (0.974, 0.981) | 47,255 | 610 |
| Extremely dense | 0.964 (0.932, 0.973) | 9398 | 90 |
| Unkown | 0.970 (0.955, 0.983) | 2069 | 53 |
| Manufacturer | |||
| GE | 0.984 (0.968, 0.993) | 5708 | 47 |
| Medison | 0.990 (0.974, 0.996) | 2673 | 13 |
| Philips | 0.977 (0.970, 0.982) | 28,943 | 412 |
| Siemens | 0.974 (0.968, 0.980) | 37,572 | 699 |
| Toshiba | 0.986 (0.978, 0.992) | 4180 | 77 |
| Other | — | 2 | 0 |