Large Language Models as Unified Multimodal Learners for Clinical Prediction

Explain Like I'm Normal

Researchers discovered that converting complex medical data—like vitals, lab results, and doctor notes—into a single text string outperforms specialized multi-model architectures. By treating everything as a language problem, a standard LLM achieved state-of-the-art results in predicting mortality and graft failure without task-specific engineering. This simplifies clinical AI by proving that 'just use an LLM' is often better than building custom fusion models.