MedDP : A Graph‐Based Large Language Disease Prediction Model

ABSTRACT

Disease prediction from electronic health records (EHRs) is challenged by noisy clinical data, limited labelled data, and the lack of interpretable evidence supporting predictions. We propose MedDP, a graph‐based disease prediction framework that integrates large language models (LLMs) with structured diagnostic knowledge. MedDP extracts and normalises clinical entities from patient records restricted to information available before the prediction time, and constructs patient‐specific heterogeneous diagnostic graphs. A graph neural network encodes these graphs and retrieves candidate diseases, which are then assessed using LLM‐generated reasoning chains scored for semantic coherence and graph consistency. MedDP also incorporates pseudo‐label‐based self‐training to leverage unlabeled data in low‐resource settings. We evaluate MedDP on MIMIC‐III and DDXPlus. On MIMIC‐III, MedDP achieves an F1‐score of 0.6724 and an area under the precision–recall curve (AUPRC) of 0.812, outperforming the other baselines reported in the paper. On DDXPlus, it achieves an average F1‐score of 0.8163 across three diagnostic granularity levels. In few‐shot experiments on a 20,000‐patient MIMIC‐III subdataset, MedDP obtains an F1‐score of 0.512 using only 1% labelled data. These results indicate that combining graph‐based diagnostic representations with LLM reasoning can support disease prediction in the evaluated datasets and settings.