Maternal Health Risk Prediction

A machine learning research project that predicts high-risk pregnancy outcomes, including preeclampsia, by combining graph representation learning with classical and deep classifiers. Patient clinical data — blood pressure, blood sugar, BMI, biomarkers — is used to build patient similarity graphs, from which node embeddings are learned and passed to downstream classifiers.

Clinical risk depends on how a patient compares to similar patients, not just on individual readings. The goal was to capture that relational structure while keeping predictions interpretable enough to trust.

1. 01

   Patient similarity graphs

   Built graphs from combined maternal health and preeclampsia datasets, connecting clinically similar patients.

2. 02

   Graph embeddings

   Learned node embeddings with multiple architectures — GCN, GraphSAGE, and GNN variants.

3. 03

   Hybrid models

   Compared embeddings across XGBoost, LightGBM, Random Forest, MLP, and LSTM; the main pipeline pairs a GCN (with reconstruction and domain-adaptation layers) with a tuned XGBoost classifier.

4. 04

   Rigour & interpretability

   Hyperparameter tuning and cross-validation, with SHAP and t-SNE for interpretability, plus synthetic-data experiments.

• A systematic comparison of graph-based hybrid models for maternal health risk.
• An interpretable pipeline with SHAP and t-SNE analysis.

“In healthcare ML, an explanation is part of the result. SHAP made the model's behaviour something we could actually discuss.”

Need something similar? See my AI & Automation service.