Adaptation Interfaces for In-Context Tabular Foundation Models in Time-to-Event Prediction

The researchers evaluated the tabular base model (TabFM) adaptation interface in time-event predictions on 74 single-risk datasets and 4 competitive-risk datasets. The study revised the context resampling training process and compared three methods: zero-shot reasoning, classification-based fine-tuning, and survival head adaptation. The results showed that CoxPH performed most reliably on the integrated Breyer score (IBS) of larger datasets; DeepHit was superior to IBS in terms of time-dependent consistency index; and causal specificity multitask learning (MTLR) ranked highest in competitive risk analysis. Zero-shot reasoning was effective on small single-risk datasets, while supervised adaptation became more advantageous as the dataset size increased; although classification fine-tuning became more competitive with increasing size, it remained weaker in probabilistic prediction.