SurvICL
A foundation model for survival analysis that casts survival analysis as an in-context learning problem. It is pretrained on synthetic data and, at inference time, predicts a full survival curve directly from a few samples without gradient updates, surpassing the traditional Cox proportional hazards model.
Timeline (last 90 days)
SurvICL is the first survival analysis model to rank above the Cox model on small datasets.
SurvICL turns survival analysis into an in-context learning problem: at inference time the model predicts the full survival curve directly from a small number of input samples, without relying on gradient updates to adapt to a new dataset.
SurvICL is pretrained entirely on synthetic data, using a flexible proportional odds prior to generate synthetic datasets, and ablation experiments show that this prior outperforms other candidate schemes in generation quality.
SurvICL is trained with an ordinary quantile regression loss, because the synthetic data generator naturally provides uncensored event times, thereby moving the hard problem of censoring handling forward into the data generation stage.
The SurvICL team built a survival analysis benchmark covering 103 datasets, the largest survival analysis benchmark to date.
On 63 held-out datasets that took no part in any design decision, SurvICL was compared against 20 baseline models, with no tuning whatsoever.
On the 63 datasets, SurvICL wins 82.9% of pairwise comparisons using the integrated Brier score, while the second-ranked Cox model has a win rate of 71.3%.
SurvICL's findings suggest that when the data characteristics of the target task do not match standard supervised learning, starting from the pretraining data generation mechanism may be more effective than modifying the model architecture or the loss function.
All Facts (8)
SurvICL is the first survival analysis model to rank above the Cox model on small datasets.
50%UnverifiedSurvICL turns survival analysis into an in-context learning problem: at inference time the model predicts the full survival curve directly from a small number of input samples, without relying on gradient updates to adapt to a new dataset.
50%UnverifiedSurvICL is pretrained entirely on synthetic data, using a flexible proportional odds prior to generate synthetic datasets, and ablation experiments show that this prior outperforms other candidate schemes in generation quality.
50%UnverifiedSurvICL is trained with an ordinary quantile regression loss, because the synthetic data generator naturally provides uncensored event times, thereby moving the hard problem of censoring handling forward into the data generation stage.
50%UnverifiedThe SurvICL team built a survival analysis benchmark covering 103 datasets, the largest survival analysis benchmark to date.
50%UnverifiedOn 63 held-out datasets that took no part in any design decision, SurvICL was compared against 20 baseline models, with no tuning whatsoever.
50%UnverifiedOn the 63 datasets, SurvICL wins 82.9% of pairwise comparisons using the integrated Brier score, while the second-ranked Cox model has a win rate of 71.3%.
50%UnverifiedSurvICL's findings suggest that when the data characteristics of the target task do not match standard supervised learning, starting from the pretraining data generation mechanism may be more effective than modifying the model architecture or the loss function.
50%