A survival model predicts event curves from censored data
Cette page n’a pas encore de traduction vérifiée. La version anglaise est affichée. Lire la version anglaise
En termes simples
Survival analysis estimates how likely an event is to remain in the future. Sometimes observation ends first, so we know only that the event will happen later than the last recorded time.
SurvICL is a model that learns from these partial records and predicts a survival curve for each person in a new dataset, using that dataset’s examples without further training.
On small held-out datasets, it ranked above the Cox model by one measure, but the difference was not statistically significant.
La recherche
When observation ends before an event occurs, the record gives only a lower bound on its timing. Can a model use these partial records to predict survival probabilities, especially in small datasets where tested alternatives have not outperformed the Cox model?
SurvICL learns from censored examples in context and predicts continuous survival curves for new datasets without dataset-specific tuning.
The authors compared SurvICL with 20 methods on 63 datasets reserved for final evaluation, using ten event-stratified train-test splits per dataset. The integrated Brier score measures prediction error for survival probabilities, with lower scores better.
A win rate counts dataset-and-competitor pairs where a method scored lower, not individuals or predictions.
SurvICL won 82.9 percent of these pairs, versus 71.3 percent for Cox, the runner-up. This aggregate comparison supports an advantage across the tested pairs, while a small-dataset subset showed a higher win rate than Cox that was not statistically significant.
The strongest result is an aggregate across the paper’s 63 held-out datasets and does not mean SurvICL wins on every dataset or against every method. SurvivalBoost had the best integrated Brier score on datasets with more than 1,000 samples, and SurvICL addresses right-censored, single-event data.
Analyse
Détails
- Auteurs de l’article original
- Julie Alberge, Linus Bleistein, Judith Abécassis, Marine Le Morvan, Gaël Varoquaux
- Soumis à arXiv
- Licence
- CC BY 4.0