Survival analysis, a statistical method for predicting the time until an event occurs, has historically faced challenges due to censored data and small clinical cohorts. While tabular foundation models (TFMs) have shown strong performance in classification and regression, they have not consistently surpassed established survival models like the Cox model in predicting time-to-event outcomes.
A new model, SurvICL (Survival In-Context), aims to bridge this gap. SurvICL is a tabular foundation model that predicts survival curves through in-context learning. Unlike many existing survival models, SurvICL is trained using a quantile regression loss, a method made possible by its reliance on synthetic pre-training data that provides uncensored event times. This approach allows the model to directly model cumulative failure probabilities, which has been found to be more stable than classical methods that use binary classifiers for discrete-time hazards.
The development of SurvICL involved pretraining the model on purely synthetic datasets. These datasets are generated from a flexible proportional-odds prior, which outperformed other prior types in ablation studies. This synthetic data generation framework allows explicit control over covariates and time-event distributions, which is particularly beneficial for survival analysis where real-world labeled data can be scarce and subject to privacy restrictions.
The researchers behind SurvICL have also introduced a new, larger survival benchmark for evaluation. Their evaluation across 53 real-world datasets showed that SurvICL, utilizing its classification formulation, outperforms both classical and deep learning baselines across various survival metrics. The model produces individualized survival predictions in a single forward pass, eliminating the need for task-specific training or hyperparameter tuning.
Other recent research has also explored adapting tabular foundation models for survival analysis. One framework reformulates survival analysis as a series of binary classification problems, handling censored observations as examples with missing labels. This allows existing TFMs to perform survival analysis through in-context learning without explicit training. Another approach involves a lightweight adaptation method for clinical survival analysis, where a "survival-aware head" is trained on top of pretrained tabular foundation model representations. This method, evaluated on large-scale ICU cohorts, showed competitive or superior performance compared to strong baselines. These findings collectively suggest that combining pretrained tabular representations with survival-aware objectives offers a practical and effective alternative for clinical survival prediction.