Comparative Analysis of Pre-Trained Language Models for Medical Document Classification and Priority-Based Workflow Routing

Abstract

Medical document processing in healthcare systems faces significant challenges due to the exponential growth in data volume and the complexity of clinical terminology. This paper presents a comprehensive comparative analysis of pre-trained language models for medical document classification and priority-based workflow routing. We evaluate BioBERT, ClinicalBERT, and base BERT models through systematic fine-tuning on diverse medical document types, including clinical notes, diagnostic reports, and insurance claims. Our multi-task learning architecture simultaneously performs document classification and priority scoring, achieving 94.7% classification accuracy and an AUC-ROC of 0.928 for urgency detection. The proposed approach reduces per-document handling time by 99.9%, cutting average manual review from 4.3 minutes per document to 0.31 seconds, while maintaining high accuracy across heterogeneous medical texts. Experimental results on 45,000 annotated medical documents demonstrate that domain-adapted models outperform general-purpose transformers by 8.3 percentage points. The integration of shared representation learning with task-specific output layers enables efficient workflow optimization, allowing for the processing of documents at 0.31 seconds per item with GPU acceleration. These findings provide actionable insights for healthcare organizations implementing automated document management systems.