Application of Textual Representation Methods for Clinical Numerical Data in Early Sepsis Diagnosis

Sepsis is a severe infectious disease with high incidence and mortality rates worldwide. Early diagnosis of sepsis in newly admitted intensive care unit patients is crucial to reduce mortality and improve patient outcomes. The manual diagnostic methods heavily rely on subjective clinical experience, while traditional machine learning methods require time-consuming feature engineering and the performance is limited by the knowledge acquired from scarce datasets. Therefore, to address the aforementioned issues, this study proposes a novel textual representation method for clinical numerical data, leveraging pre-trained language models from the field of natural language processing for sepsis prediction. Specifically, this study innovatively transforms structured clinical numerical data of patients into unstructured textual descriptions. This transformation reframes sepsis prediction into a text classification task, leveraging the rich prior semantic knowledge embedded in pre-trained language models to enhance prediction performance. The proposed method is validated using real ICU clinical data. When employing RoBERTa-base, it achieved an F1 score of 79.03%, which represents an improvement of five percentage points compared with commonly used machine learning classifiers. The experiments confirmed that the proposed method enhances the performance of early sepsis diagnosis and introduces new insights for clinical diagnosis of sepsis.