AI and the World's Hard Problems · Medicine & drug discovery · Nature (computational biology) · 2026-10-22
Large language models for multimodal clinical predictive modeling: a systematic review
Electronic health records (EHR) contain both structured and unstructured data, and the emergence of large language models has enabled unstructured modalities to be incorporated into predictive models built on EHR data. Of 641 studies screened, 57 met the inclusion criteria, and the review identified three LLM-based multimodal integration strategies: embedding, extraction, and text fusion. All strategies yielded performance gains over single-modality models, with an average 6.4% increase, and the authors outline considerations including sample size, interpretability, and computational costs.