Capabilities of LLMs and Few-Shot Learning: A study on Cardiac MRI Data

Clinical decision making is often based on analysis of retrospective patient cohorts to enable and guide treatment decisions. In general, there are large volumes of data within hospitals within Electronic Patient Record systems as unstructured data and some of the challenges in extracting relevant information involves manual resource from the limited number of expert annotators with sufficient domain knowledge.

In this study, we explored the use of Large Language Models and prompting-based techniques to extract information from CMR reports that contain patient information and multiple measurements, which otherwise would require manual extraction and transcription to databases without errors.

We have also evuated on training customised models in a few-shot setting when minimal annotated data is available and computational resources such as GPUs are not. Our study evaluates the adaptability and performance of LLMs on our hospital data can provide useful insights for applications in other real-world settings.