Table 2 Overall performance of the Clinical Transformer compared with other modeling approaches. Results are reported as mean c-index ± standard deviation across the 10 train/test splits
Modeling framework | Learning strategy | MYSTIC trial (NSCLC) | OAK trial (NSCLC) | Samstein et al.33 pan-cancer | Thorsson et al.34 pan-cancer (TCGA) | Chowell et al.3 pan-cancer | Chowell et al.3 evaluated on MYSTIC |
|---|---|---|---|---|---|---|---|
Clinical Transformer | Neural network | 0.670 ± 0.07a | 0.669 ± 0.04 | 0.649 ± 0.02b | 0.734 ± 0.01 | 0.720 ± 0.01 | 0.643 ± 0.004a |
Cox-nnet | Neural network | 0.609 ± 0.04 | 0.626 ± 0.04 | 0.622 ± 0.02 | 0.676 ± 0.01 | 0.707 ± 0.01 | c |
DeepSurv | Neural network | 0.602 ± 0.05 | 0.620 ± 0.04 | 0.595 ± 0.02 | 0.707 ± 0.01 | 0.691 ± 0.01 | c |
Neural MTLR | Neural network | 0.626 ± 0.05 | 0.577 ± 0.03 | 0.538 ± 0.02 | 0.658 ± 0.05 | 0.680 ± 0.01 | c |
Nnet-survival | Neural network | 0.603 ± 0.04 | 0.563 ± 0.03 | 0.527 ± 0.02 | 0.691 ± 0.01 | 0.695 ± 0.02 | c |
Transformer Survival | Neural network | 0.584 ± 0.03 | 0.583 ± 0.03 | 0.608 ± 0.02 | 0.706 ± 0.006 | 0.709 ± 0.01 | c |
Linear modeling | CoxPH regression | 0.599 ± 0.03 | 0.620 ± 0.03 | 0.594 ± 0.03 | 0.690 ± 0.01 | 0.709 ± 0.01 | c |
Nonlinear modeling | Random survival forest | 0.606 ± 0.04 | 0.664 ± 0.05 | 0.638 ± 0.01 | 0.722 ± 0.01 | 0.714 ± 0.01 | c |
Biomarkers | TMB | 0.589 ± 0.05 | 0.543 ± 0.05 | 0.538 ± 0.02 | 0.624 ± 0.01 | 0.550 ± 0.02 | c |
