RonaldKato/Decoding-Maternal-Care

A Graph Neural Network Approach for Intent Classification and Linguistic Analysis of Expert-Authored Maternal Health Dialogue in Low-Resource Uganda # PrenatalGNN **A GNN + LLM knowledge-graph framework for AI-assisted prenatal SMS triage in Uganda.** > Every year, thousands of preventable maternal deaths in Uganda trace back > to one root cause: a danger sign reported by SMS that reached a clinician > hours too late. PrenatalGNN is a research framework for closing that gap > — combining a knowledge-graph-augmented GNN classifier with clinical > validation, fairness auditing, and Luganda-language support, so that > "I'm bleeding and in pain" gets triaged in minutes, not hours. ## Why this project exists Maternal health chatbots and USSD/SMS triage lines are increasingly common across East Africa, but most publish only headline accuracy numbers. This project asks the harder questions a clinical deployment actually needs answered: - **Does the AI change outcomes**, not just classify messages correctly? (→ propensity-score-matched causal analysis) - **Is it faster than the status quo**, and by how much for the cases that matter most? (→ AI-vs-human triage simulation) - **Is it fair** across Uganda's regions, not just accurate on average? (→ per-region fairness audit with equalized odds / disparate impact) - **Does it survive contact with real SMS**, i.e. noisy text and code-switching into Luganda? (→ noise-robustness and zero-shot multilingual evaluation) - **What does the medical knowledge graph actually contribute**, versus a model with no domain grounding? (→ ablation study) PrenatalGNN packages all five analyses into one reproducible pipeline, with every stage runnable standalone. ## Pipeline overview The framework implements the six-stage pipeline above end to end: 1. **Input Data** — load a real de-identified SMS/USSD export, or generate a realistic synthetic dataset (10 intent classes, clinical risk levels, 10% Luganda code-switching) with zero setup. 2. **Data Preprocessing** — column inference/mapping, encoding auto-detection, train/test split, standardization. 3. **Exploratory Data Analysi …