A Neuro-Symbolic Method for Understanding Free-text Medical Evidence
Video
Team Information
Team Members
Tian Kang, PhD Candidate, Department of Biomedical Informatics, Graduate School of Arts and Science
Ali Turfah, MA Candidate, Department of Biomedical Informatics, Graduate School of Arts and Science
Jaehyun Kim, Research Scientist, Department of Biomedical Informatics, Graduate School of Arts and Science
Adler Perotte, Assistant Professor, Department of Biomedical Informatics, Graduate School of Arts and Science
Faculty Advisor: Chunhua Weng, Professor of Biomedical Informatics, Columbia University
Abstract
Introduction: As medical evidence expands exponentially, it is increasingly hard for clinicians to practice Evidence-based Medicine due to the challenges in literature comprehension and synthesis at the point of care. Early methods reply on ontologies and lexico-syntactic patterns to extract biomedical concepts from the literature and identify candidate answers. Machine Reading Comprehension (MRC) is a promising solution. Its recent advances are dominantly enabled by deep neural networks, which, however, achieve only modest predictive gain in the biomedical domain and have limited reasoning capabilities compared to symbolic systems. We introduce Medical evidence Dependency (MD)-informed Self-Attention, a Neuro-Symbolic Model for understanding free-text medical evidence in literature. We hypothesize this method can get the best of both: the high capacity of neural networks and the rigor, semantic clarity and reusability of symbolic logic.
Methods: We develop a symbolic compositional representation called Medical evidence Dependency (MD) to represent the basic medical evidence entities and relations following the PICO framework widely adopted among clinicians for searching evidence. We use Transformer as the backbone and train one head in the Multi-Head Self-Attention to attend to MD and to pass linguistic and domain knowledge onto later layers (MD-informed). We integrated MD-informed Attention into BioBERT and evaluated it on two public MRC benchmarks for medical evidence from literature: i.e., Evidence Inference 2.0 and PubMedQA.
Results: The integration of MD-informed Attention head improves BioBERT substantially for both benchmarks—as large as by +30% in the F1 score—and achieves the new state-of-the-art performance on the Evidence Inference 2.0. By visualizing the weights learned from MD-informed Attention head, we find the model can capture clinically meaningful relations separated by long passages of text.
Conclusion: MD-informed Attention is a novel neuro-symbolic computational framework that enhances neural reading comprehension models with reusable, human-readable symbolic knowledge for better interpretability and reasoning capability. Its compositionality can benefit any Transformer-based architecture for machine reading comprehension of free-text medical evidence.
Contact this Team
Contact: Tian Kang (use form to send email)
