Extracting Non-Taxonomic and Ternary Relations From Patient-Generated Texts for Semantic Interoperability

Why It Matters

By converting unstructured patient narratives into interoperable knowledge, the framework boosts the reliability of health data analytics and supports more informed clinical decisions.

Key Takeaways

• Framework extracts non‑taxonomic and ternary relations from patient texts
• Achieves 98.91% accuracy and 77.6% F1 score
• Validation rate reaches 92.7% for 384 semantic relations
• Uses delayed fusion, rule‑based dictionary, and BioBERT
• Improves semantic interoperability for clinical decision support

Pulse Analysis

Patient‑generated texts—forums, diaries, and social media posts—contain rich clinical signals but lack the structured hierarchy required for traditional health databases. Most natural‑language processing pipelines focus on taxonomic, hierarchical relationships, leaving associative, causal, and ternary connections underexplored. This blind spot hampers the creation of comprehensive knowledge graphs, limiting the ability of clinicians and researchers to capture the nuanced interplay of symptoms, treatments, and risk factors that emerge in real‑world patient narratives.

The presented framework adopts a Design Science Research methodology and a pragmatic philosophy to bridge this gap. Its four‑layered architecture integrates contextual neural learning with an interpretable rule‑based dictionary, while a delayed fusion strategy harmonizes these components. BioBERT serves as a domain‑specific validator, ensuring factual grounding. Tested on over 38,000 anxiety and depression documents, the system extracted 222 unique concepts in a union mode, achieved 98.91% accuracy, and improved F1 by 10% versus BiLSTM baselines. Moreover, it validated 384 semantic relations—240 ternary and 144 non‑taxonomic—with a 92.7% success rate.

The implications extend beyond academic metrics. By translating free‑form patient language into a structured, interoperable knowledge graph, the framework supports more precise clinical decision support tools, facilitates cross‑domain ontology integration, and accelerates biomedical knowledge reuse. Healthcare providers can now query patient‑reported outcomes with greater semantic depth, while pharmaceutical and AI firms gain a richer data substrate for drug safety monitoring and predictive modeling. As regulatory bodies push for enhanced data interoperability, solutions that capture non‑hierarchical relations will become essential components of next‑generation health informatics ecosystems.