Prediction Performance and Cif Validity in Transformer-Based Models for Competing Risks

censoring and competing risks. Censoring happens when a subject does not experience any event in the duration of the study. Competing risks meanthat several mutually exclusive events can occur. Classical models such as Cox and Fine-Gray models are widely used in survival analysis, but theirproportional cause-specific and subdistribution hazards assumption can be restrictive. In this paper, we investigate transformer-based models for survivalanalysis on the SEER dataset. We compare prediction-horizon temporal and static feature transformers paired with several output heads againstclassical baselines and a multilayer perceptron (MLP) paired with the same output heads using Brier score, integrated Brier score and C-index. We evaluate whether the predicted cumulative incidence functions (CIFs) are non-decreasing over time and whether their sum across all causes does not exceed one. We also evaluate a post-hoc projection as a correction tool for structurally invalid predictions. Our results show that the temporal transformer did not materially outperform the MLP baseline on static SEER covariates, whereas the static feature transformer showed small but consistent gains. Across backbones, endpoint performance was similar among output heads, but CIF validity depended on output-head construction. The hazard and increment CIF heads produced valid CIFs by construction, while the direct CIF head produced violations of monotonicity and the constraint that the sum of cause-specific CIFs must not exceed one. Dykstra projection removed these violations with negligible changes in endpoint metrics and should be interpreted as a structural correction rather than a calibration improvement. This shows the need to evaluate both predictive performance andCIF shape when using neural models in survival analysis.