Decoding Maternal Care: A Graph Neural Network Approach for Intent Classification and Linguistic Analysis of Expert Authored Maternal Health Dialogue in the Global South

Expert clinical communication in the Global South encodes a kind of tacit knowledge that most maternal-health NLP work never touches: how trained professionals decide what to say first, how much to say, and in what tone, under real constraints on time, text length, and access. We study this using the MOTHER dataset, 503 clinician-authored maternal-health question-answer pairs from Uganda, combining domain-adapted preprocessing, six-topic LDA, VADER-based sentiment analysis with clinical lexical re-weighting, and an intent-entity co-occurrence knowledge graph, on top of which we train an inductive GraphSAGE classifier to recover clinical intent across ten categories. The topic structure that emerges (NPMI = 0.61) tracks the antenatal-care continuum itself rather than arbitrary lexical clusters, and the fact that 92.2% of responses score neutral-to-positive in tone quantifies something clinically meaningful: experts default to reassurance even when discussing risk, a pattern any downstream dialogue system would need to preserve rather than flatten. This relational organization turns out to matter for classification, not just for description: the graph-based model reaches a macro-F1 of 0.914 (10-fold CV std. 0.03), well above a Node2Vec+MLP graph-embedding baseline (0.79) and standard TF-IDF classifiers (0.62-0.68), and an ablation isolating the graph structure and domain lexicon shows both are jointly responsible for this margin rather than either alone -- evidence that what experts are doing is genuinely relational, linking symptoms, reassurance, and follow-up through shared clinical entities rather than shared vocabulary. This same graph, used to ground a retrieval-augmented generation module, produces responses judged less prone to hallucination than two general-purpose medical LLMs under expert review, suggesting that the structure experts implicitly impose on their own knowledge is a more efficient scaffold for a dialogue system than raw model scale. Together, these results indicate that expert dialogue from the Global South does not need to be treated as noisy data to be normalized away: its structure is itself the signal, and modeling it explicitly lets a small, interpretable, auditable system approach the communicative judgment of trained clinicians in a domain where transparency matters as much as accuracy.