Breaching the curation bottleneck with human-machine reading symbiosis

PurposeThe explosion of molecular biomarker and treatment information in the precision medicine era drastically exacerbated difficulty in identifying patient-relevant knowledge for clinical researchers and practitioners. Curated knowledgebases, such as the JAX Clinical Knowledgebase (CKB) are tools to organize and display knowledge in a readily accessible format; however, curators face the same challenges in comprehensively identifying clinically relevant information for curation. Natural language processing (NLP) has emerged as a promising direction for accelerating manual curation, but prior applications were often conceived as stand-alone efforts to automate curation, and the scope is often limited to simple entity and relation extraction. In this paper, we study the alternative paradigm of assisted curation and identify key desiderata to scale up knowledge curation with human-computer symbiosis. MethodsWe chose precision oncology for a case study and introduced self-supervised machine reading, which can automatically generate noisy training examples from unlabeled text. We developed a curation user interface (UI) for precision oncology and through iterative "curathons" (curation hackathons), conducted retrospective and prospective user studies for head-to-head comparison between manual and machine-assisted curation. ResultsContrary to the prevailing assumption, we showed that high recall is more important for end-to-end assisted curation. In extensive user studies, we showed that assisted curation can double the curation speed and increase the number of findings by an order of magnitude for previously scarcely curated drugs. ConclusionWe demonstrated that an iterative and thoughtful collaboration between professional curators and NLP researchers can facilitate rapid advances in assisted curation for precision medicine. Human-machine reading symbiosis can potentially be applicable to clinical care and research scenarios where curation is a major bottleneck.