Tabular Foundation Models Can Do Survival Analysis
Organizations: Imperial College London
Abstract
While tabular foundation models have achieved remarkable success in classification and regression, adapting them to model time-to-event outcomes for survival analysis is non-trivial due to right-censoring, where data observations may end before the event of interest occurs. In this work, we systematically study a variety of approaches to utilize off-the-shelf tabular foundation models originally designed for classification for discrete-time survival modeling. We evaluate existing classification formulations, including hazard modeling and multi-class classification approaches, and introduce a new formulation that directly models cumulative failure probabilities at discretized time points, treating censored observations as examples with missing labels at certain time points. We prove that under standard censoring assumptions, minimizing the weighted cross entropy loss recovers the true survival probabilities as the training set size increases. Across real-world datasets ( static and dynamic), we find that TFMs paired with classification reductions often outperform or are competitive with classical and deep learning survival baselines on average across multiple survival metrics. Our results show that off-the-shelf TFMs provide a flexible approach to survival modeling, allowing time discretization to be adapted at inference without training. Our comparison provides practical guidance on when to use these models and which survival formulation to choose.
Figures & tables
| Model | C-Index | Integrated AUC | Integrated Brier Score | Avg. Rank | |||
|---|---|---|---|---|---|---|---|
| Value | Rank | Value | Rank | Value | Rank | ||
| Cumulative-Failure Models | |||||||
| MITRA-CF | 0.699 ±0.016 | 6.3 ±0.4 | 0.752 ±0.016 | 5.6 ±0.4 | 0.152 ±0.007 | 4.7 ±0.3 | 5.5 ±0.3 |
| TabPFN-CF | 0.665 ±0.018 | 8.9 ±0.6 | 0.727 ±0.017 | 9.2 ±0.5 | 0.158 ±0.007 | 6.7 ±0.4 | 8.2 ±0.5 |
| XGBoost-CF | 0.688 ±0.015 | 8.0 ±0.5 | 0.737 ±0.016 | 7.5 ±0.5 | 0.156 ±0.007 | 5.4 ±0.5 | 7.0 ±0.4 |
| Discrete-time Hazard Models | |||||||
| Model | C-Index | Integrated AUC | Integrated Brier Score | Avg. Rank | |||
|---|---|---|---|---|---|---|---|
| Value | Rank | Value | Rank | Value | Rank | ||
| Classification-based Models (CF, ours) | |||||||
| MITRA-CF | 0.674 ±0.049 | 6.5 ±1.5 | 0.655 ±0.031 | 5.8 ±1.3 | 0.162 ±0.032 | 6.7 ±1.6 | 6.3 ±1.3 |
| TabPFN-CF | 0.636 ±0.064 | 5.3 ±1.4 | 0.642 ±0.051 | 5.4 ±1.5 | 0.164 ±0.029 | 8.0 ±0.8 | 6.3 ±0.9 |
| XGBoost-CF | 0.588 ±0.028 | 8.0 ±1.1 | 0.588 ±0.041 | 7.5 ±1.6 | 0.158 ±0.029 | 6.8 ±1.3 | 7.4 ±1.2 |
| Discrete-time Hazard Variants (DH) | |||||||
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| No. | Dataset | Covariates | Samples | Censor Rate |
|---|---|---|---|---|
| 1 | actg | 11 | 1151 | 91.7 |
| 2 | Aids2 | 4 | 2839 | 38.0 |
| 3 | Bergamaschi | 10 | 82 | 65.9 |
| 4 | breast | 4 | 100 | 74.0 |
| 5 | burn | 11 | 154 | 68.8 |
| 6 | cancer | 8 | 228 | 27.6 |
| Dataset | Total Samples | Training Samples | Covariates | (i) | (ii) | (iii) |
|---|---|---|---|---|---|---|
| acath | 3504 | 2452 | 3 | ✓ | ||
| AML_Bull | 116 | 80 | 6283 | ✓ | ||
| chop | 414 | 289 | 3833 | ✓ | ||
| DBCD | 295 | 205 | 4919 | ✓ | ||
| DLBCL | 240 | 167 | 7399 | ✓ | ||
| gse1992 | 124 | 86 | 15541 | ✓ |
| No. | Dataset | No. of covariates | No. of Samples | Censoring Rate | |
|---|---|---|---|---|---|
| 1 | aids | 5 | 467 | 59.7 | 6 |
| 2 | csl | 6 | 446 | 39.5 | 8 |
| 3 | epileptic | 5 | 556 | 66.2 | 7 |
| 4 | heart | 4 | 103 | 27.2 | 5 |
| 5 | oldmort | 7 | 4603 | 57.2 | 10 |
| 6 | FRTCS | 5 | 697 | 89.7 | 8 |
| C-Index | Integrated AUC | IBS | ||||||||||
| Model | ||||||||||||
| Cumulative-Failure Models | ||||||||||||
| MITRA-CF | 0.694 | 0.699 | 0.702 | 0.702 | 0.758 | 0.751 | 0.752 | 0.748 | 0.155 | 0.152 | 0.150 | 0.150 |
| TabPFN-CF | 0.672 | 0.660 | 0.663 | 0.665 | 0.739 | 0.721 | 0.726 | 0.720 | 0.157 | 0.160 | 0.158 | 0.157 |
| XGBoost-CF | 0.682 | 0.687 | 0.695 | 0.688 | 0.738 | 0.732 | 0.740 | 0.737 | 0.159 | 0.158 | 0.154 | 0.153 |
| Discrete-time Hazard Models | ||||||||||||
| Before Transformation | C-Index | Integrated AUC | IBS | ||||||
|---|---|---|---|---|---|---|---|---|---|
| Model | Violations | Size | Clipping | Isotonic | Clipping | Isotonic | Clipping | Isotonic | |
| MITRA-CF | 5 | 1.5% | 0.003 | 0.694 | 0.694 | 0.758 | 0.758 | 0.155 | 0.155 |
| 10 | 3.7% | 0.007 | 0.699 | 0.699 | 0.751 | 0.751 | 0.152 | 0.152 | |
| 15 | 5.7% | 0.008 | 0.700 | 0.700 | 0.749 | 0.749 | 0.151 | 0.151 | |
| 20 | 9.1% | 0.005 | 0.700 | 0.700 | 0.745 | 0.745 | 0.151 | 0.151 | |
| All | 5.0% | 0.006 | 0.699 | 0.699 | 0.752 | 0.752 | 0.152 | 0.152 | |
| Mean | |||
|---|---|---|---|
| Sampling Strategy | C-Index | AUC | IBS |
| Uniform | 0.688 | 0.743 | 0.134 |
| Sample-stratified | 0.689 | 0.740 | 0.134 |
| Horizon-stratified | 0.687 | 0.742 | 0.133 |
| Datasets where the TFM is better (%) | Relative gain (%) | |||||
|---|---|---|---|---|---|---|
| Comparison | C-Index | Integrated AUC | IBS | C-Index | Integrated AUC | IBS |
| MITRA-CF vs. XGBoost-CF | ||||||
| TabPFN-CF vs. XGBoost-CF | ||||||
| MITRA-DH vs. XGBoost-DH | ||||||
| TabPFN-DH vs. XGBoost-DH | ||||||
Survival Analysis
Survival analysis models when events such as disease progression, failure, or death occur, accounting for incomplete follow-up and censoring.
See all 112 papers