Deep Contextual Transfer Learning Models for Clinical Information Extraction

Abstract

Clinical Information Extraction aims to extract meaningful information from clinical text to support various healthcare applications including decision support, pharmacovigilance, clinical research. Our research investigates deep neural models and novel transfer learning strategies to push the state-of-the-art in clinical information extraction, with a specific emphasis on infrequently occurring low-resource entities such as negation, speculation, and adverse drug events. Our research contributes generic methodologies to enhance the transferability of con temporary biomedical contextual language models in a computationally inexpensive fashion, potentially benefiting various biomedical natural language processing (NLP) applications. We explore the systematic synthesis of multi-task learning and neural transfer learning to improve neural model generalizability and reduce neural model redundancy. Taken together, we present contextual multi-task learning , a generic framework for enhanced information extraction with specific emphasis on clinical decision support and pharmacovigilance. newline