Abstract
Predicting patient outcomes from clinical notes is crucial for proactive healthcare management. This paper introduces HERMES (Healthcare Reasoning via Medical Entity Semantics), a novel framework that leverages contrast-aware knowledge graph reasoning to predict patient outcomes from unstructured clinical notes. Our approach constructs patient-specific knowledge graphs from clinical narratives and employs contrast learning to identify discriminative patterns between different outcome groups. By explicitly modeling the relationships between medical entities and incorporating contrastive signals, HERMES achieves superior prediction performance while providing interpretable reasoning paths. Experimental results on real-world clinical datasets demonstrate significant improvements over existing methods.
Abstract
Predicting patient outcomes from clinical notes is crucial for proactive healthcare management. This paper introduces HERMES (Healthcare Reasoning via Medical Entity Semantics), a novel framework that leverages contrast-aware knowledge graph reasoning to predict patient outcomes from unstructured clinical notes. Our approach constructs patient-specific knowledge graphs from clinical narratives and employs contrast learning to identify discriminative patterns between different outcome groups. By explicitly modeling the relationships between medical entities and incorporating contrastive signals, HERMES achieves superior prediction performance while providing interpretable reasoning paths. Experimental results on real-world clinical datasets demonstrate significant improvements over existing methods.
Key Contributions
- Contrast-Aware Reasoning: Novel integration of contrastive learning with knowledge graph reasoning
- Clinical Note Processing: Effective extraction of structured knowledge from unstructured text
- Patient-Specific KGs: Construction of personalized knowledge graphs from clinical narratives
- Interpretable Predictions: Transparent reasoning paths for clinical decision support
- Superior Performance: Significant improvements in patient outcome prediction accuracy
Method Overview
HERMES Framework
- Knowledge Graph Construction:
- Extraction of medical entities from clinical notes
- Relationship modeling between entities
- Patient-specific graph generation
- Contrast-Aware Reasoning:
- Contrastive learning to identify discriminative patterns
- Comparison of similar patients with different outcomes
- Enhanced feature representation through contrast
- Outcome Prediction:
- Graph-based reasoning for prediction
- Attention mechanism over reasoning paths
- Interpretable decision making
Conference Details
- Conference: CITA 2026 (The 15th Conference on Information Technology and its Applications)
- Status: Under Review
- Year: 2025
Keywords
Clinical Notes, Knowledge Graphs, Patient Outcome Prediction, Contrastive Learning, Healthcare AI, Medical NLP