Towards Semi-Automated Construction of Laboratory Test Result Comprehension Knowledgebase for a Patient-Facing Application

AbstractViewing lab test results is the most frequent activity patients do when accessing patient portals but lab results can be confusing for patients. There is a critical need to build tools to help patients better comprehend lab results and formulate questions to prepare for physician consults. In this project, we collected and annotated 251 textual records of lab results comprehension from AHealthyMe.com. To scale up the data collection and curation with automated named entity recognition, we evaluated different transformer-based language models including BioBERT, ClinicalBERT, RoBERTa_trial (RoBERTa model further pre-trained with clinical trial eligibility criteria corpus ofClinicalTrials.gov), and PubMedBERT. Results showed that RoBERTa_trial is the best performing model based on strict criteria and PubMedBERT model has the best performance based on lenient criteria. The models are expected to serve as an AI feature for the design and development of patient-facing technology called LabGenie in promoting lab results communication between providers and patients.