Neuro-Symbolic AI for Women's Health.

Area of Science:

• Artificial Intelligence in Medicine
• Biomedical Informatics
• Computational Linguistics

Background:

• Women's health issues like menopause and infertility affect a significant global population.
• The vast and growing biomedical literature presents challenges for current AI in extracting reliable insights.
• Actionable, machine-interpretable representations of diseases are crucial for biomedical research.

Purpose of the Study:

• To investigate the extent to which ontologies and knowledge graphs (symbolic AI) can support human-centric explainable AI for artificial neural networks (neural AI).
• To explore a neuro-symbolic AI approach combining neural and symbolic AI for processing biomedical text.
• To assess the impact of incorporating domain knowledge into AI models for enhanced customization and explainability.

Main Methods:

• A neuro-symbolic AI approach was developed, integrating neural AI for pattern extraction with symbolic AI for background knowledge representation.
• Domain knowledge and scientific evidence from biomedical literature were leveraged to create human-readable explanations (explainable AI) as nanopublications (knowledge graphs).
• Experiments involved applying unsupervised vector arithmetic formulas (cosine, 3CosAdd) to word2vec embeddings from PubMed citations and evaluating Large Language Models (LLMs) for term extraction.

Main Results:

• Three experiments (EXP1-EXP3) evaluated 315 n-grams from word2vec embeddings derived from over 300,000 PubMed citations.
• A fourth experiment (EXP4) assessed 381 terms extracted by 9 LLMs (including specialized biomedical and general models) from evidence-based text.
• The study explored the utility of incorporating prior domain knowledge into vector arithmetic for customizing AI model outputs.

Conclusions:

• Biomedical knowledge can guide the explainability of neural model predictions, identifying which predictions warrant explanation and which can be disregarded.
• Integrating biomedical knowledge enhances the customizability of AI models, particularly when using word2vec embeddings and LLMs for health issue pattern extraction.
• The neuro-symbolic approach offers a pathway to more reliable and interpretable AI insights from the biomedical literature.