Table 2 Comparison of survival prediction methods on TCGA-BRCA and TCGA-BLCA datasets
From: HONeYBEE: enabling scalable multimodal AI in oncology through foundation model-driven embeddings
Category | Method | BRCA | BLCA |
|---|---|---|---|
Baseline methods | SurvPath29 | 0.655 ± 0.089 | 0.625 ± 0.056 |
|  | ABMIL (KP)30 | 0.615 ± 0.083 | 0.566 ± 0.038 |
|  | MCAT31 | 0.652 ± 0.117 | 0.598 ± 0.094 |
|  | MOTCat32 | 0.600 ± 0.095 | 0.596 ± 0.079 |
|  | Porpoise33 | 0.652 ± 0.042 | 0.636 ± 0.024 |
|  | PathOmic34 | – | 0.586 ± 0.062 |
Individual modalities | |||
Clinical features | Cox | 0.920 ± 0.017 | 0.823 ± 0.011 |
|  | RSF | 0.845 ± 0.020 | 0.799 ± 0.026 |
|  | DeepSurv | 0.928 ± 0.013 | 0.842 ± 0.018 |
Pathology report features | Cox | 0.495 ± 0.056 | 0.556 ± 0.044 |
|  | RSF | 0.511 ± 0.066 | 0.534 ± 0.038 |
|  | DeepSurv | 0.497 ± 0.047 | 0.561 ± 0.025 |
Molecular features | Cox | 0.525 ± 0.066 | 0.501 ± 0.026 |
|  | RSF | 0.426 ± 0.061 | 0.536 ± 0.045 |
|  | DeepSurv | 0.520 ± 0.012 | 0.468 ± 0.015 |
WSI features | Cox | 0.461 ± 0.051 | 0.519 ± 0.053 |
|  | RSF | 0.491 ± 0.049 | 0.524 ± 0.036 |
|  | DeepSurv | 0.464 ± 0.051 | 0.517 ± 0.048 |
Multimodal fusion | |||
Concatenation | Cox | 0.767 ± 0.052 | 0.799 ± 0.022 |
|  | RSF | 0.808 ± 0.029 | 0.790 ± 0.028 |
|  | DeepSurv | 0.847 ± 0.040 | 0.820 ± 0.015 |
Mean pooling | Cox | 0.589 ± 0.042 | 0.648 ± 0.041 |
|  | RSF | 0.592 ± 0.083 | 0.577 ± 0.061 |
|  | DeepSurv | 0.666 ± 0.048 | 0.689 ± 0.019 |
Kronecker product | Cox | 0.728 ± 0.030 | 0.709 ± 0.055 |
|  | RSF | 0.625 ± 0.068 | 0.705 ± 0.028 |
|  | DeepSurv | 0.758 ± 0.023 | 0.721 ± 0.018 |
