BSO-AD: An Ontology for Representing and Harmonizing Behavioral Social Knowledge in ADRD

The authors developed BSOAD, the first ontology to systematically represent and harmonize behavioral and social factors influencing Alzheimer's disease and related dementias by integrating existing ontologies and literature-derived relationships, while validating its structure and coverage through expert review and an LLM-assisted evaluation framework.

The Gist

Imagine you are trying to solve a massive, complex puzzle: Alzheimer's disease and related dementias (ADRD).

For a long time, scientists have been looking at the puzzle pieces in two very different rooms.

• Room A (The Biology Room): Contains pieces about genes, brain chemicals, and drugs.
• Room B (The Life Room): Contains pieces about a person's life, like their job, their education, whether they have friends, if they can afford healthy food, or if they live alone.

The problem is that these two rooms don't talk to each other. The scientists in Room A speak a different language than the scientists in Room B. Because of this, it's hard to see the whole picture of how a person's life (like stress or loneliness) actually changes their brain biology to cause dementia.

Enter BSO-AD: The Universal Translator and Puzzle Master.

This paper introduces a new digital tool called BSO-AD. Think of it as a giant, super-smart digital filing cabinet or a universal translator that finally connects Room A and Room B.

How does it work? (The Creative Analogy)

Imagine you are trying to organize a chaotic library where books are written in 50 different languages and stored in random piles.

1. The Problem: You have a book about "Smoking" written in medical code, and another about "Loneliness" written in social science jargon. They are both important for understanding dementia, but they don't fit together.
2. The Solution (BSO-AD): The researchers built a new, organized library system.
   • They didn't start from scratch. They took the best shelves from existing libraries (like the Social Determinants of Health library and the Drug Repurposing library) and glued them together.
   • They added a special "Rosetta Stone" (using standard medical codes like ICD-10) so that a doctor's note saying "Patient is unemployed" instantly matches the library's category for "Economic Instability."
   • They created bridges between the shelves. They didn't just list "Smoking" and "Dementia" separately; they built a bridge labeled "Smoking increases the risk of Dementia." They even built bridges that show how it happens (e.g., Smoking → Damages Blood Vessels → Affects Brain → Dementia).

What did they build?

• 2,275 Categories: They created a massive map with over 2,000 specific categories, from "What you eat" to "Your neighborhood safety" to "Specific brain tests."
• The "Why" and "How": They didn't just list facts. They mapped out the relationships.
  • Direct Link: "Lack of education" is linked to "Higher risk of dementia."
  • Indirect Link: "Lack of education" might lead to "Lower income," which leads to "Poor nutrition," which affects "Brain inflammation," which leads to "Dementia."
• The "AI Librarian" Test: To make sure their new library was actually good, they didn't just ask humans to read it (though they did that too). They used Artificial Intelligence (AI) as a super-fast librarian.
  • The AI read thousands of scientific papers.
  • It asked: "Does this library have a place for every concept I found in these papers?"
  • The Result: The library was excellent! It covered 97% of the concepts the AI found, and the connections between the shelves made perfect logical sense.

Why does this matter to you?

Think of BSO-AD as the foundation for a new kind of detective work.

Before this, if a doctor wanted to know how a patient's social life affects their dementia, they had to guess or look at messy, disconnected notes. Now, with BSO-AD:

• Computers can finally "read" the whole story. AI can scan millions of patient records and scientific papers, using this map to find hidden patterns.
• Better Prevention: We might discover that fixing a specific social issue (like helping a senior find a community center) is just as powerful as a new drug in slowing down dementia.
• Personalized Care: It helps doctors see the whole patient, not just their brain scan.

The Bottom Line

The researchers built a universal map that connects how we live (our behaviors and social world) with how our brains age. By using this map, we can finally stop treating the "social" and "medical" sides of dementia as separate problems and start solving them together. It's like finally putting all the puzzle pieces on one table so we can see the full picture.

Technical Summary

1. Problem Statement

Alzheimer's Disease and Related Dementias (ADRD) are a growing public health crisis heavily influenced by Behavioral and Social Factors (BSFs), such as physical activity, social isolation, and socioeconomic status. However, integrating BSF knowledge into ADRD research faces significant challenges:

• Data Heterogeneity: BSF data is scattered across unstructured sources (clinical narratives, literature) and structured sources (EHRs, surveys) using inconsistent standards.
• Semantic Silos: Existing ontologies are fragmented. Biological/clinical ontologies (e.g., ADO, DROADO) focus on molecular mechanisms, while social ontologies (e.g., SDoHO, OMRSE) focus on determinants of health in isolation.
• Lack of Integration: There is no unified framework to systematically represent the interplay between BSFs and ADRD, hindering scalable evidence synthesis, computational analysis, and the development of targeted interventions.

2. Methodology

The authors developed BSO-AD (Behavioral Social Data and Knowledge Ontology for ADRD) following established biomedical ontology engineering principles (OBO Foundry, FAIR principles).

• Development Strategy:

  • Modular Reuse: The ontology was built by reusing and extending existing ontologies to ensure semantic interoperability:
    • SDoHO (Social Determinants of Health Ontology): For core BSF concepts.
    • DROADO (Drug-Repurposing Oriented ADRD Ontology): For molecular and pharmacological mechanisms.
    • AD-Onto: For neuropsychological assessments (e.g., MMSE subitems).
    • TEO (Time Event Ontology): For temporal dimensions.
  • Standardization: Integrated ICD-10-CM Z codes (Z55–Z65) to map socioeconomic and psychosocial factors to clinical coding systems, addressing a gap in SDoHO.
  • Hybrid Approach: Used a top-down approach for broad domain concepts and a bottom-up approach to integrate specific classes from literature and ICD codes.

• Relationship Modeling:

  • Direct Associations: Mapped BSFs to ADRD using a hierarchy derived from the UMLS Semantic Network. Relationships include functional (e.g., affectsRiskOf, causes, predicts) and temporal (e.g., associatedWithShorterTime) links.
  • Indirect/Mechanistic Associations: Modeled multi-level pathways where BSFs influence intermediate biological entities (genes, pathways, pathologies), which in turn affect ADRD.

• Evaluation Framework:

  • Expert Review: Used Hootation to convert ontology axioms into natural language sentences for manual validation by domain experts.
  • LLM-Assisted Automated Evaluation: Developed a novel pipeline using Large Language Models (Llama4, Qwen3, MedGemma) to:
    1. Construct a domain-specific literature corpus.
    2. Extract and classify BSF concepts.
    3. Calculate Category Coverage (how well the ontology covers literature concepts).
    4. Compute Semantic Coherence metrics: Completeness (overlap with literature), Conciseness (specificity of ontology classes), Child Similarity Score (CSS), and Parent-Child Similarity Score (PSS) using ClinicalBERT embeddings.

3. Key Contributions

1. First Integrated Ontology: BSO-AD is the first ontology to systematically formalize BSFs specifically within the context of ADRD, bridging the gap between social/behavioral science and clinical neurology.
2. Multi-Layer Relationship Framework: It introduces a sophisticated relational structure that captures both direct associations and indirect mechanistic pathways (via genes/pathways), enabling multi-scale knowledge representation.
3. LLM-Assisted Evaluation: The study proposes and validates a scalable, automated framework for ontology assessment using LLMs and embedding-based metrics, moving beyond traditional manual review to handle large-scale semantic validation.
4. Interoperability: By integrating ICD-10-CM codes and reusing established ontologies, BSO-AD is designed for direct application in EHR-based knowledge integration and AI-driven analysis.

4. Results

• Ontology Statistics: BSO-AD comprises 2,275 classes, 153 object properties, and 49 data properties, with over 3,000 logical axioms.
• Expert Review: Achieved a 0.95 rational agreement and 0.96 inter-evaluator agreement. Disagreements (e.g., incorrect parent-child relationships like "Non-smoker is a type of Smoker") were resolved through discussion, leading to structural refinements.
• Automated Evaluation Metrics:
  • Category Coverage: Extremely high across all BSF domains (≥ 0.97), indicating the ontology successfully captures the breadth of concepts found in literature.
  • Completeness: Average of 0.79, showing strong alignment between ontology classes and literature-derived concepts.
  • Conciseness: Average of 0.94, indicating that ontology classes are specific and relevant to the domain.
  • Semantic Coherence: CSS and PSS scores were stable (0.71–0.80), confirming a coherent hierarchical structure. The "Food" category showed lower scores, identifying it as an area needing structural enrichment (e.g., adding intermediate classes).

5. Significance

• Foundation for AI-Driven Research: BSO-AD provides a semantic backbone for integrating heterogeneous data sources (EHRs, literature, mobile apps), enabling advanced computational modeling of how social factors drive ADRD.
• Precision Medicine & Policy: By formalizing the link between BSFs and ADRD outcomes, the ontology supports the development of targeted prevention strategies and health policies tailored to specific risk factors.
• Methodological Advancement: The paper demonstrates the feasibility of using LLMs for automated ontology evaluation, offering a scalable solution for maintaining and assessing complex biomedical knowledge graphs.
• Future Impact: The framework is designed to evolve, with plans to incorporate ICD-11 codes and expand via large-scale data mining, positioning it as a critical infrastructure for future ADRD research.