JMIR Publications

ABSTRACT

Background:

Sensor metadata is critical for exposure health research because it supports accurate sensor identification, deployments, data integration, interoperability, and reproducibility. Yet it is often fragmented across multiple heterogeneous sources, such as scientific literature and manufacturer guides, where key specifications are frequently reported indirectly through citation chains, making reference tracing essential for metadata enrichment and completeness.

Objective:

To address this bottleneck, we developed and evaluated an LLM-based automated, citation-aware pipeline that enriches sensor metadata extracted from a primary article by identifying sensor-related citation markers and extracting additional metadata from the referenced sources.

Methods:

We extend our prior LLM-based metadata extraction approach by (i) detecting sensor mentions in full-text articles, (ii) capturing nearby citation markers, (iii) resolving markers to full bibliographic entries in the reference list, and (iv) retrieving cited papers to extract additional sensor metadata that may be absent from the primary document and using it to enrich and complete the base metadata.

Results:

Across 20 primary papers, the citation extraction component achieved 74.2% precision, 92.0% recall, 82.1% F1-score, and 69.7% accuracy, and all extracted bibliographic entries were correctly matched to their source references. This component increased sensor extraction by about 261%, yielding 94 additional sensors overall.

Conclusions:

The developed citation-guided pipeline improved sensor discovery and metadata completeness, thereby supporting the development of richer, more complete sensor metadata repositories.