Sienna Mao
Research Mentor(s): Yun Jiang
Mentor Department:
Authors: Xinyue Mao, Yun Jiang
Session: Session 3 (11:00am – 11:50am)
Presentation Type: Poster 71
Abstract
Patients with cancer often face a high risk of hospital readmission, which can lead to poor patient outcomes and increased medical burden. It is crucial to be proactive and identify high-risk individuals with precise prediction models and enable timely interventions. Traditional models typically rely on structured data including sociodemographic characteristics, comorbidity records, and clinical lab results. However, unstructured data, such as clinical notes, also contain valuable insights that may enhance the predictive accuracy of patients’ readmission. This study, using EHR data from MIMIC-IV, explored how incorporating information extracted from clinical notes using Large Language Models (LLMs) could improve the prediction capabilities of 30-day rehospitalization and identified key factors influencing readmission in patients diagnosed with breast, colorectal, lung, or prostate cancer. Baseline machine learning and deep learning models using only structured data were compared with models integrating features extracted from clinical notes. The integrated models demonstrated better performances, achieving an AUC of 0.70 for breast cancer, 0.65 for colorectal cancer, 0.67 for lung cancer, and 0.68 for prostate cancer.
