Table 4:
Model Classification of n-year lung cancer risk across selected cohorts
| NLST-test (n=898) | NLST-testnodules (n=896) | LS-A (n=882) | LI-A (n=219) | Consortium-A (n=364) | Consortium-B (n=117) | Consortium-DECAMP (n=131) | Consortium-C (n=115) | BRONCH (n=373) | Average Rank (range) n=‰ | ||
|---|---|---|---|---|---|---|---|---|---|---|---|
| Input | Model | ||||||||||
| Clinical variables | Mayo | ‡ | 0.804 [0.798, 0.809] | ‡ | ‡ | 0.706 [0.704, 0.708] | 0.864 [0.862, 0.867] | 0.568 [0.565, 0.571] | 0.716 [0.712, 0.719] | 0.621 [0.615, 0.628] | 3.5 (7, 1) n=6 |
| Clinical variables | Brock | ‡ | 0.789 [0.782, 0.796] | ‡ | ‡ | 0.716 [0.714, 0.718] | 0.885 [0.883, 0.886] | 0.662 [0.659, 0.666] | 0.713 [0.710, 0.716] | 0.497 [0.494, 0.499] | 3.2 (5, 2) n=6 |
| Single CT AI | Liao et al. | 0.751 [0.747, 0.756] | 0.755 [0.750, 0.759] | 0.723 [0.712, 0.734] | 0.644 [0.635, 0.653] | 0.662 [0.660, 0.664] | 0.779 [0.776, 0.782] | 0.706 [0.703, 0.709] | 0.660 [0.656, 0.663] | 0.621 [0.614, 0.628] | 3.9 (6, 1) n=9 |
| Single CT AI | Sybil | 0.881 [0.877, 0.885] * | 0.879 [0.872, 0.885] * | 0.779 [0.768, 0.789] | 0.763 [0.756, 0.770] | 0.700 [0.694, 0.706] | 0.889 [0.884, 0.895] | 0.606 [0.597, 0.616] | 0.764 [0.756, 0.772] | 0.623 [0.618, 0.629] | 2.6 (6, 1) n=9 |
| Longitudi nal CT AI | DLSTM | 0.738 [0.734, 0.743] | 0.727 [0.721, 0.731] | ¶ | 0.711 [0.702, 0.720] | 0.743 [0.741, 0.745] | § | 0.778 [0.774, 0.781] | § | § | 3.6 (6, 2) n=5 |
| Longitudi nal CT AI | TDViT | 0.797 [0.793, 0.802] | 0.790 [0.785, 0.794] | ¶ | 0.773 [0.764, 0.781] * | 0.753 [0.750, 0.755] | § | 0.823 [0.820, 0.825] * | § | § | 1.8 (3, 1) n=5 |
| Multimodal | DLS | 0.783 [0.778, 0.788] | 0.776 [0.771, 0.782] | 0.810 [0.799, 0.820] * | † | † | † | † | † | † | 2.7 (4, 1) n=3 |
| Multimodal | DLI | † | † | † | † | 0.856 [0.854, 0.858] * | 0.936 [0.935, 0.938] * | 0.742 [0.739, 0.745] | 0.851 [0.849, 0.854] * | † | 1.5 (3, 1) n=4 |
Note.—Data are reported as bootstrapped mean area under the receiver operating characteristic curve [95% CI]. The n-year lung cancer risk for each cohort was 2-year risk for each cohort except LI-A, which was 3-year risk, and BRONCH, which was 1-year risk.
result was significantly different compared to each other method in the column for p<0.01
nodule characteristics unavailable (missing >10% of nodule size, attenuation, count, spiculation, or lobe location)
prohibitive class imbalance (only 6/23 lung cancer cases have more than one scan)
Missing demographic, smoking history, COPD, or CYFRA covariates
No longitudinal imaging
‰ Number of cohort evaluations performed with this model