همایش بین المللی هوش مصنوعی

صفحه اصلی / دومین همایش بین المللی هوش مصنوعی

Active Learning Strategies for Imbalanced Clinical Note Categories in ICU Mortality Prediction

نویسندگان :

Mina Afzali¹ Mohsen Kahani² Maryam Karrabi³ Nadieh Armin⁴

1- دانشگاه فردوسی مشهد 2- دانشگاه فردوسی مشهد 3- دانشگاه فردوسی مشهد 4- دانشگاه فردوسی مشهد

کلمات کلیدی :

mortality prediction،unstructured clinical notes،data imbalance،active learning،Note embedding،MIMIC III

چکیده :

Accurate mortality prediction in intensive care units (ICUs) is critical for improving patient outcomes and optimizing clinical decision-making. Clinical notes contain rich unstructured information that can complement structured data in predictive models. However, these datasets are often highly imbalanced, with far fewer mortality cases than survivals, limiting the performance of traditional machine learning methods. In this study, we perform a comparative analysis of different clinical note categories for ICU mortality prediction and introduce an active learning framework to address data imbalance. By iteratively selecting the most informative samples for model training, active learning improves predictive accuracy while reducing annotation costs, resulting in faster model training and effective management of data imbalance. Our experiments demonstrate that the application of Active Learning led to a significant reduction in both model execution time and the number of labeled samples required for training. Crucially, this enhancement in efficiency was concurrent with an improvement in predictive performance, as quantified by elevated AUC scores for several categories.