Combining generative large language models and pretrained language models for predicting acute respiratory distress syndrome progression.

Acute Respiratory Distress Syndrome (ARDS) is a life-threatening condition in which early diagnosis and timely intervention are crucial. Recent research into machine learning for predicting ARDS progression based on objective clinical data has been hindered by the limited availability of high-quality annotated datasets. Pretrained Language Models (PLMs) have shown promising results in various Natural Language Processing (NLP) tasks, especially in unstructured electronic health records, but their application to numerical clinical data has been constrained by data type limitations. This study investigates the feasibility of combining generative Large Language Models (LLMs) with PLMs for predicting ARDS progression. The proposed method leverages LLMs to convert numerical clinical indicators into clinical narratives, followed by classification using PLMs to predict ARDS progression. The proposed method was evaluated using two prominent public ICU databases (eICU and MIMIC-IV) and a real-world dataset from the First Affiliated Hospital of Anhui Medical University. Validation results indicate that our method surpasses traditional approaches, yielding F1-scores of 0.8128, 0.9333, and 0.8462, respectively. These results show that generative LLMs and PLMs can improve early ARDS prediction and clinical decisions.