MediPhen: Prompt-Based LLM Reasoning with Synthesized Multimodal Clinical Knowledge for Zero-Shot Multi-morbidity Phenotyping

Clinical decision support using heterogeneous electronic health records (EHRs) is a well-established yet rapidly expanding research area. Large language model (LLM)-driven approaches have shown dominant performance in processing unstructured data such as clinical notes for disease phenotype classification. However, the absence of a unified reasoning framework capable of integrating structured laboratory results with unstructured clinical notes under zero-shot conditions limits progress in multimodal clinical decision support. To address this gap, we propose MediPhen, a novel reasoning framework that transfers LLMs for multi-morbidity disease phenotyping using multimodal clinical data. MediPhen introduces a framework for adapting LLMs to zero-shot disease phenotyping by incorporating extracted clinical entities, their relations, and lab narratives from EHRs, integrating a clinical knowledgebase to guide phenotype classification and enhance LLM transfer learning performance, and an explanation module that leverages chain-of-thought prompting to improve clinical reasoning. Comprehensive experiments conducted on MIMIC-III and MIMIC-IV benchmarks across multiple LLMs demonstrate the effectiveness of MediPhen. Notably, MedGemma-27B achieved state-of-the-art performance, improving micro averaged F1 scores by 19.92% on MIMIC-III and 16.68% on MIMIC-IV compared to fine-tuned baselines. These results highlight MediPhen as a zero-shot screening tool for multi morbidity phenotype classification, scalable within research infrastructures, advancing integration of structured and unstructured EHR data in clinical AI.