An iterative knowledge extraction and dual-channel fusion framework for clinical differential diagnosis

For common diseases in the cardiovascular, cerebrovascular, and respiratory systems that present with overlapping clinical manifestations (for instance, chest tightness and shortness of breath), accurate identification is critical for formulating effective treatment plans. However, the unstructured nature of medical records and the implicit relationships embedded in medical knowledge pose significant challenges to automated diagnosis. Existing methods overlook medical commonsense and triplet relational structures, rendering them incapable of capturing symptom–disease associations and consequently limiting their discriminative ability for diseases with similar clinical presentations. To address these limitations, we construct the Chinese Clinical Document–Electronic Medical Record (CCD-EMR) dataset for complex common diseases and propose the Knowledge-enhanced Iterative Diagnostic Framework (KIDF). This model achieves in-depth analysis of complex clinical phenotypes through a knowledge extraction–dual-channel fusion mechanism. Specifically, the medical knowledge triplet extraction module combines large language models with a rule-based knowledge base to extract high-fidelity structured knowledge from free-text medical records. The dual-channel fusion module integrates BERT and TransE to fuse medical record semantics with graph embeddings, which are subsequently refined by a hybrid feature classifier to accurately capture local key patterns and long-range semantic dependencies. Experimental results demonstrate that the KIDF model achieves 77.19% accuracy and a 62.94% macro-average F1 score on the CCD-EMR dataset. Ablation studies further confirm that the proposed modules play a crucial role in enhancing diagnostic performance for complex common diseases.