Efficiency of foundation models on detecting patient outcomes from clinical data

Survival analysis plays a central role in clinical research, where predicting the time until an event such as death is essential to support decision-making. While classical statistical and machine learning models are well established for this task, the recent emergence of tabular foundation models - large models pretrained on many datasets and applied through in-context learning - raises the question of whether they can also be useful for survival analysis on clinical data. This thesis investigates the use of tabular foundation models as feature extractors for clinical survival analysis. A pipeline is developed in which two pretrained models, TabPFN and TabICL, produce embeddings from the input data, and a set of survival heads - a Cox model, a Random Survival Forest, and neural DeepSurv models - is trained on top of these embeddings. The same heads applied to the raw features serve as baselines, so that the contribution of the embeddings can be isolated. The approach is evaluated on two real clinical datasets using the Antolini time-dependent concordance index, and is studied with respect to model comparison, training set size, robustness to missing values, and interpretability through SHAP. The results show that the embedding-based approach is a viable alternative to classical models but does not generally outperform them, with TabPFN consistently stronger than TabICL. Its main advantage emerges in specific settings, most notably in the low-data regime, where the TabPFN embeddings clearly outperform the baselines. The SHAP analysis further reveals that the choice of backbone, rather than the survival head, primarily determines which clinical features drive the predictions.