Authors - Thinh Truong, Chau Vo, Anh Duong Abstract - The early diagnosis of patient conditions at the hospital admission stage is crucial for optimizing medical resource allocation, reducing overcrowding, and improving patient outcomes. Traditional diagnostic approaches at admission rely on limited initial information and expert assessment, which can lead to misclassification and delayed treatment. This paper proposes a multimodal data-driven approach that integrates Large Language Model (LLM) to predict patient conditions using structured and unstructured medical data. In particular, we propose a classification model that leverages LLM for multimodal data processing and generates feature representation based on demographics, biometrics, vital signs, lab values and electrocardiogram (ECG) data for 78-disease diagnoses. Compared to the existing models, our model decides a better data fusion with semantics-preserving. Indeed, evaluated through experiments on the constructed dataset from MIMIC-IV using standard metrics such as Area Under the Receiver Operating Characteristic (AUROC), Precision, Recall, and F1-score, the proposed model outperforms traditional ones. Experimental results also highlight the potential of integrating multiple data sources for automated patient triage at the admission stage.
