Neuro-Symbolic AI for Women's Health

BACKGROUND: Menopause, endometriosis, miscarriage, and female infertility are health issues affecting women worldwide (nearly half the global population). The biomedical literature is human-readable and evergrowing with around 3.5K papers published daily. Current Artificial Intelligence (AI) cannot reliably deal with facts, and automatic processing of scientific publications to obtain reliable insights remains a challenge, although, representing diseases in an actionable (machine-interpretable semantics) has a long-standing tradition in biomedical research. OBJECTIVE: Conduct some experiments to explore to what extent ontologies and knowledge graphs (symbolic AI) can support comprehensive human-centric explainable AI for artificial neural networks (neural AI). METHODS: Instead of using Large Language Models (LLMs) from neural AI as all-in-one generative AI solution, this paper investigates a neuro-symbolic AI approach, combining neural AI (to process and extract patterns for health issues from free-text) with symbolic AI (explicit representations of background knowledge). Our neuro-symbolic AI approach leverages on domain knowledge (simple classification based on predefined categories), and scientific evidence from the biomedical literature to provide human-readable explanations (explainable AI) formally represented as nanopublications (machine-processable knowledge graphs). We investigated if incorporating prior domain knowledge (best scientific evidence) into vector arithmetic formulas may support customisation (e.g. bringing "unseeing" terms for a predefined category). RESULTS: We performed three experiments (EXP1, EXP2 and EXP3), evaluating 315 candidate n-grams obtained by applying unsupervised vector arithmetic formulas (cosine for similarity and 3CosAdd for four-term analogies) to word2vec embeddings created from 301,201 PubMed citations (titles and abstracts). We also conducted a fourth experiment (EXP4) with 9 LLMs to automatically extract and classify terms from evidence-based text excerpts, evaluating 381 terms from LLMs' output. In EXP4, we looked into the output categories provided by 2 open-source small-size biomedical LLMs (with 66.4 and 184 millions of trainable parameters) and 7 free-of-charge general LLMs (DeepSeek-V3, Groq, Grok3-beta, QWEN2.5-MAX, Gemini, Claude, ChatGPT4). CONCLUSION: Biomedical knowledge may guide explainability (what predictions from neural models are worthy to explain and what predictions can be ignored) and enable a higher level of customisation when using word2vec embeddings and LLMs for extracting patterns for health issues from free-text.