Multimodal Survival Analysis with Locally Deployable Language Models

Our research explores how artificial intelligence can improve medical prognosis while protecting patient privacy. We developed a lightweight, locally deployable large language model (LLM) that predicts patient survival by combining information from clinical text, medical records, and genetic data.

Unlike traditional AI systems that depend on cloud computing, our model runs entirely on-site, making it suitable for hospitals with limited computing resources and strict data-security requirements. It not only provides accurate, calibrated survival predictions, but also generates concise, human-readable explanations that help clinicians understand the reasoning behind each estimate.

Tested on real cancer patient data, our approach outperforms existing survival models and demonstrates that trustworthy, interpretable AI can be achieved without compromising privacy or accessibility.

Lead Investigator:

• Christopher Yau