Medical data is inherently messy, spanning unstructured clinical notes, high-resolution pathology images, and complex genomic sequences. The fundamental challenge in medical AI is bridging the "semantic gap"—the disconnect between low-level data points (like pixels or word frequencies) and high-level clinical abstractions (like a diagnosis or a treatment plan).
To bridge this gap, researchers have converged on hierarchical knowledge representation. By organizing medical concepts into structured trees or graphs, systems can move beyond simple pattern matching to perform reasoning that mimics human clinical logic.
Bridging the Semantic Gap
The transition from raw data to actionable insight requires a structured abstraction process. As early as 1980, Okada argued in Conceptual Taxonomy of Japanese Verbs for Uderstanding Natural Language and Picture Patterns that a hierarchical system is necessary to link visual patterns with linguistic concepts. In medicine, this manifests as a multi-layered organization of knowledge.
Dasmahapatra et al. formalize this in Living with the Semantic Gap: Experiences and Remedies in the Context of Medical Imaging by proposing modular ontologies. These structures organize knowledge both vertically (by imaging modality) and horizontally (by granularity), allowing a system to map low-level visual features to high-level expert interpretations. Similarly, Lou et al. describe a "PathoGraph" in PathoGraph: A Graph-Based Method for Standardized Representation of Pathology Knowledge that layers information from cellular morphology up to complex diagnostic logic.
Ontologies as the Semantic Backbone
If raw data is the fuel, ontologies are the engine. Ebadollahi et al. argue in Concept-based electronic health records: opportunities and challenges that ontologies serve as the essential semantic backbone for health records, transforming them from simple document storage into decision-enabling information.
These hierarchies do more than just categorize; they define the relationships that allow models to understand context. For example:
- Prioritization: Feder et al. found that clinical-focused ontologies, which prioritize hierarchical links between conditions, are far more effective for clinical summarization than billing-focused ones (Building a Clinically-Focused Problem List From Medical Notes).
- Semantic Enrichment: Yue et al. demonstrate in Med-Tree: A Medical Ontology Tree Combined with the Graph Attention Networks for Medication Recommendation that structuring medical entities in a tree captures semantic relationships that purely data-driven methods miss, leading to safer prescriptions.
Differentiating Overlapping Symptoms
One of the most difficult tasks in medicine is differential diagnosis—distinguishing between two diseases that present with similar symptoms. A flat representation of medical literature struggles here because the symptoms appear identical in text.
Hierarchical structures solve this by clustering conditions based on shared manifestations. The DIALOG system, introduced in 1975, used an "attention focusing heuristic" within a hierarchy to partition diagnostic hypotheses. Modern iterations, such as the MedRAG framework proposed by Zhao et al. in A Smart Multimodal Healthcare Copilot with Powerful LLM Reasoning, use hierarchical knowledge graphs to identify "unique feature triplets." By navigating from broad categories to specific tiers, the model can pinpoint the exact manifestation that distinguishes one disease from another.
Knowledge Transfer and Sparse Data
Hierarchies also provide a mechanism for "knowledge transfer." In many medical domains, data for specific rare diseases is sparse. However, by using graph-based attention on medical ontologies, models can leverage the parent-child relationships inherent in disease classification.
As the authors of TrialBench: Multi-Modal AI-Ready Datasets for Clinical Trial Prediction (Chen et al.) explain, embedding these hierarchical structures allows a model to transfer knowledge from well-documented parent concepts to data-poor sub-categories. This structured expertise is even being distilled into large-scale models; Zhou et al. developed a "Pathology Knowledge Tree" to pretrain models that can identify tumors in zero-shot scenarios, relying on the hierarchy rather than explicit training labels (Knowledge-Enhanced Visual-Language Pretraining for Computational Pathology).
By organizing medical knowledge into these multi-dimensional, hierarchical structures, researchers are moving toward what Jin et al. call a "holographic" patient portrait—a comprehensive, structured representation that captures the full complexity of human health (Research on the Construction and Application of Breast Cancer-Specific Database System Based on Full Data Lifecycle).
Key Foundational Work
DIALOG: A Model Of Diagnostic Logic For Internal Medicine (1975) — contributed 2 concept elements to this direction.
Conceptual Taxonomy of Japanese Verbs for Uderstanding Natural Language and Picture Patterns (1980) — contributed 2 concept elements to this direction.