Supporting clinical decisions utilizing machine learning and radiology notes.

In recent years the use of large language models (LLMs) in the clinical domain has enabled clinical decision support through the use of electronic medical records (EMRs). While these advancements have opened up the way for many complex tasks such as report coding, sentence classification, and summarization their impact on prediction tasks has been overshadowed. In this study, the ability of LLMs is evaluated with traditional processing techniques and models in binary prediction tasks of coronary atherosclerosis diagnosis and patient mortality using MIMIC IV radiology reports. It finds that XGBoost on a bag of words (BOW) dataset is competitive with a pre-trained RadBERT model.

Medical staff makes decisions that impact patient health every day using data generated from a variety of sources. While tabular and image formats can be extremely useful, the written narrative format is the easiest way to distribute complex information. This information, while actionable, cannot be used by computer systems to provide decision support without large amounts of preprocessing. In this experiment, radiology notes are evaluated in the methods of processing to provide patient outcome predictions on binary diagnosis and mortality. The results show traditional methods of XGBoost and a bag of words are comparable with large language models in their ability to predict patient outcomes.

Team Members

Advisors