Extracting Clinical Relationships from Discharge Summaries of Supra Sellar Lesion Patients using Gemini LLM

A suprasellar lesion refers to an abnormal growth or mass that is located above the saddle-shaped depression at the base of the skull, known as sella turcica. The proposed work utilizes the textual content of the discussion section of 553 discharge summaries of suprasellar lesion patients, collected from NIMHANS Hospitals, Bangalore. A rich variety of clinical information is contained in these clinical texts. As there is a lack of annotated data, this study attempts to extract the 3 kinds of relations between the clinical concepts in an unsupervised manner using Gemini, a Large Language Model. The three types of relations between the clinical entities (Problem, Treatment and Test) are Treatment Administered for Problem ( TrAP ), Test Reveals Problem ( TeRP ) and Problem indicates Problem ( PIP ). Appropriate prompts were provided to Gemini for the clinical relation extraction. Initially, a clinical NER model in the Stanza toolkit is used to extract the clinical entities from the documents and the output is provided to Gemini using the prompt for relation extraction. This model achieved an F1-score of 0.66. Additionally, another experiment was carried out by directly giving the text from the discharge summary to Gemini for relation extraction which generated an F1-score of 0.81. This showed the ability of Gemini to identify more relations between the clinical concepts related to suprasellar lesions.