Presentation (CTI)

OBJECTIVES: Time-to-event projection involves balancing relevant clinical information with statistical fit considerations to ensure that the selected model is appropriate. We explored how large language models (LLMs) can be leveraged for many of the required tasks currently done manually.
METHODS: Understanding the prognosis of the study population is important to establish plausible ranges for life expectancy or median survival. The use of iterative prompting to a LLM to obtain information on these parameters was tested in a study requiring overall survival with levamisole-based treatment for stage B/C colon cancer.
RESULTS: The LLM was able to identify the relevant study, review it, provide a detailed summary of the risk profile of the study population, and “best estimates of the median and mean years these patients could expect to survive without access to the newer treatments”. In addition to these estimates (5-7 years), the response included a description of the reasoning, steps taken make the estimates, and factors that were considered (e.g., disease stage mix). When provided with alternative models to select from (without any contextual information on methods or guidelines), the LLM factored in statistical fit, plausibility of projection and implications in an economic model to support its recommendations. While it did not automatically reject models with implausible tails, when prompted about them, it provided suggestions for alternative approaches and strategies to mitigate the impact (e.g., capping) in an economic model.
CONCLUSIONS: With structured prompting and provision of some contextual information, LLMs can facilitate the inclusion of clinical input in time-to-event projection analyses. This could be advanced further using APIs to integrate the AI capabilities into agentic tools that streamline the analytic process. The well-known risks of hallucinations and errors remain a concern in such applications, compelling human review and active involvement of clinical experts.