Key Coverage Matters: Semi-Structured Extraction of OCR Clinical Reports

Researchers developed a semi-structured extraction method for digitizing fragmented clinical reports using OCR and question-answering models, introducing 'key coverage' as a metric to measure data completeness. The approach achieves F1 scores above 0.83 on real-world hospital data from 20+ institutions using a lightweight BERT model, demonstrating that canonical key inventory completeness drives extraction performance.

Healthcare data fragmentation across institutions represents a critical infrastructure challenge that limits patient care quality and research capabilities. This research addresses the practical problem of converting paper and scanned clinical reports into structured digital data, enabling better EHR integration and longitudinal patient analysis. The innovation lies not just in the extraction technique but in the key coverage framework—a systematic approach to building and validating canonical field inventories through iterative mining, normalization, and clustering.

The clinical records digitization challenge has grown more urgent as healthcare systems increasingly recognize the value of comprehensive patient histories for personalized medicine, drug research, and clinical trial matching. OCR technology alone proves insufficient due to document heterogeneity and noise, requiring downstream intelligence to map variable terminology to canonical fields. Prior approaches often assumed fixed field structures, but clinical documents vary significantly across hospitals and regions.

This method's practical value extends beyond academic application. Healthcare institutions face pressure to reduce costs while maintaining on-premise deployment capabilities for data sensitivity reasons. The 0.2B parameter BERT model meets these constraints while achieving competitive performance, outperforming a larger Qwen3 baseline at comparable coverage levels. The language-agnostic design proves particularly significant—the methodology successfully generalizes despite using Chinese medical records, suggesting applicability across healthcare systems globally.

The key coverage metric itself warrants industry attention as it provides measurable progress tracking for data standardization efforts. As hospitals increasingly digitize archives and seek interoperability, having a principled framework for evaluating inventory completeness enables data managers to allocate resources strategically. Future implementation will likely focus on cross-institutional key mapping and automated field discovery in new healthcare contexts.

• →Key coverage metric quantifies clinical data inventory completeness and demonstrates monotonic performance improvement with canonical field coverage.
• →Lightweight BERT-based model achieves 0.839 F1 exact-match performance on real hospital OCR documents from 20+ institutions.
• →Semi-structured extraction approach designed for low-cost on-premise deployment addresses healthcare privacy and data-silo constraints.
• →Language-agnostic methodology enables adaptation across different healthcare systems despite training on Chinese medical records.
• →Top-90 canonical keys represent practical threshold for high-performance extraction in heterogeneous clinical document sets.