Review of Language Models for Survival Analysis

By learning statistical relations between words, Large Language Models (LLMs) have presented the capacity to capture meaningful representations for tasks beyond the ones they were trained for. LLMs’ widespread accessibility and flexibility have attracted interest among medical practitioners, leading to extensive exploration of their utility in medical prognostic and diagnostic applications. Our work reviews LLMs’ use for survival analysis, a statistical tool for estimating the time to an event of interest and, consequently, medical risk. We propose a classification of LLMs’ modelling strategies and adaptations to survival analysis, detailing their limitations and strengths. Due to the absence of standardised guidelines in the literature, we introduce a framework to assess the efficacy of diverse LLM strategies for survival analysis.

Code available on GitHub.