Creating a Biomedical Causal Knowledge Graph via Semi-Supervised Learning

Technical Summary: Creating a Biomedical Causal Knowledge Graph via Semi-Supervised Learning

Problem Statement
The construction of biomedical causal knowledge graphs (CKGs) faces a critical bottleneck: the scarcity of large-scale, expert-annotated corpora required for supervised relation extraction. While biomedical knowledge graphs are essential for organizing heterogeneous data and supporting clinical reasoning, distinguishing directed cause-effect relationships from mere statistical associations requires nuanced domain expertise. Manual annotation is prohibitively expensive at the scale needed for modern deep learning, particularly given the high lexical diversity and epistemic caution (e.g., distinguishing "associated with" from "causes") inherent in biomedical literature. Consequently, there is a need for methods that can leverage the vast quantities of existing unlabeled biomedical text to overcome this labeling constraint while preserving the causal specificity required for clinical validity.

Methodology
This study employs a mixed-methods approach combining quantitative experiments with qualitative graph analysis, structured around a self-training semi-supervised learning framework.

• Data Sources: The research utilizes the BioCause corpus (19 full-text articles, 850 causal relations) as the primary labeled dataset for infectious diseases. For unlabeled data, 50 additional full-text articles from the same subdomain were collected. Evaluation references include the Drug-Disease Causal Knowledge Graph (DD-CKG) and the OFFSIDES dataset for adverse drug effects.
• Semi-Supervised Framework: The core methodology follows a self-training paradigm with specific adaptations for causal extraction:
  • Feature Extraction: Utilizes lexical (tokens, lemmas, POS), syntactic (dependency relations), and semantic (named entities, ontological categories) features derived from automatic parses and ontology lookups.
  • Base Classifier: A Support Vector Machine (SVM) with a radial basis function kernel is trained on the labeled BioCause data.
  • Conditional Thresholding: A critical innovation in this study is the use of adaptive confidence thresholds rather than fixed global thresholds. Thresholds are calculated separately for each relation type based on the empirical distribution of correct predictions. This prevents the exclusion of valid instances that fall below a rigid global cutoff while filtering out noisy predictions.
  • Iterative Refinement: The model iteratively predicts on unlabeled data, adds high-confidence instances to the training set, and retrains until convergence or performance degradation on a development set.
• Knowledge Graph Construction: Extracted relations undergo entity normalization (mapping to SNOMED CT, RxNorm, and Entrez Gene identifiers), relation aggregation (using confidence-weighted voting for duplicate edges), and confidence annotation. The resulting graph assigns reliability scores to edges based on extraction confidence and source multiplicity.
• Evaluation: Performance is measured via standard precision, recall, and F1-score on a held-out test set. Graph quality is assessed by comparing extracted drug-disease edges against the DD-CKG reference and by evaluating utility in a downstream task: predicting adverse drug effects against the OFFSIDES dataset.

Key Contributions

1. Empirical Validation of Semi-Supervised Learning: The study demonstrates that self-training with unlabeled data significantly improves causal relation extraction performance in a low-resource setting, addressing the annotation bottleneck.
2. Adaptive Thresholding Strategy: The paper introduces and validates a conditional threshold strategy for instance selection, showing it outperforms fixed confidence thresholds by accommodating the varying confidence distributions of different causal relation types.
3. Integrated Construction Pipeline: It presents a complete methodology for transforming semi-supervised extractions into a structured, confidence-annotated causal knowledge graph, including entity normalization and edge weighting.
4. Clinical Utility Demonstration: The research provides evidence that the resulting causal graph offers superior performance in downstream pharmacovigilance tasks compared to purely associative baselines.

Results

• Relation Extraction Performance: The semi-supervised model achieved an F-score of 83.47% for causal trigger detection (a 7.77 percentage point gain over the supervised baseline) and 78.2% for full relation extraction (a 9.7 percentage point gain). The improvements were most pronounced for relations with semi-explicit or less frequent triggers.
• Graph Characteristics: The constructed graph contains 12,847 entities and 34,219 causal edges. It exhibits a scale-free degree distribution with hub nodes representing central biomedical concepts (e.g., "infection," "antibiotics"). The graph includes 1,247 direct and 3,892 indirect causal paths, enabling reasoning about mediation pathways.
• Comparison with Reference: Against the DD-CKG reference, the system achieved 67.5% recall and 72.4% precision for drug-disease edges. Manual review of novel edges not in the reference indicated that approximately 60% were valid causal relationships.
• Downstream Task (Adverse Drug Effects): In predicting adverse drug reactions, the causal graph achieved an Area Under the Curve (AUC) of 0.78, compared to 0.71 for an associative baseline. Restricting the graph to high-confidence edges (score > 0.85) further improved the AUC to 0.82.

Significance and Claims
The paper claims that semi-supervised learning, specifically through self-training with adaptive thresholds, is a viable and effective strategy for scaling biomedical causal knowledge graph construction. The study argues that unlabeled biomedical text contains distributional information sufficient to learn complex causal patterns, thereby reducing reliance on expensive expert annotation.

The significance of this work lies in its demonstration that causal specificity can be preserved and enhanced through automated methods, resulting in graphs that are not only structurally coherent but also functionally useful for clinical applications. The ability to identify adverse drug effects with higher accuracy than associative methods suggests that these graphs capture information critical for pharmacovigilance and precision medicine. The author positions this approach as a necessary step toward integrating literature-derived causal knowledge with electronic health records and clinical decision support systems, while acknowledging limitations regarding corpus scale, domain generalizability, and the need for temporal maintenance of causal claims.