Citation

Moon, Md Mahmudul Hasan and Assaduzzaman, Md and Amin, Md Masrafi Al and Rahman, Md Musfiqur and Ferdous, Jannatul and Fahad, Nafiz and Liew, Tze Hui and Ohidujjaman, Ohidujjaman (2026) Explainable meta-learning framework for maternal risk prediction using clinical feature engineering. Intelligence-Based Medicine, 16. p. 100488. ISSN 2666-5212

Text 1.pdf - Published Version Restricted to Repository staff only Download (9MB)

Abstract

Early and accurate maternal health risk classification is critical for guiding timely obstetric interventions, particularly in resource-limited clinical settings. Existing approaches are hindered by unaddressed class imbalance, limited feature engineering, and insufficient model transparency. This study proposed a Stacking-Based Meta-Learning framework for binary classification of maternal health risks. The proposed pipeline integrates domain-driven feature engineering, SMOTE-Tomek hybrid class resampling, out-of-fold stacking ensemble learning, and multi-level explainable artificial intelligence (XAI) within a unified architecture. A total of 22 clinically informed features were derived from 10 physiological variables, yielding an enriched 32-variable feature representation. XGBoost and SVM served as base learners under a Logistic Regression meta-learner trained on out-of-fold probabilistic predictions. The proposed framework achieved 99.58% accuracy, an AUC-ROC of 0.9999, and an F1-score of 99.58% on the primary dataset, with a 95% confidence interval of 97.64%–99.94%, outperforming all baseline classifiers, including KNN, SVM, and XGBoost. Only one prediction was incorrect with the proposed model across the entire test set, indicating high sensitivity and specificity. To assess generalization capability, stratified 10-fold cross-validation was employed, achieving a mean accuracy of 99.43% ± 0.28%. Further evaluation with adversarial perturbations (ε = 0.01, 0.05, 0.10) and Gaussian noise confirmed that the model performs reliably even in the presence of realistic variations in clinical data quality. Global SHAP analysis, LIME-based local explanations, and uncertainty quantification provide multi-level interpretability to support clinically meaningful risk stratification. The developed framework represents a statistically robust, interpretable, and clinically applicable solution for AI-enabled maternal health risk assessment.

Item Type:	Article
Uncontrolled Keywords:	Maternal health risk prediction, Stacking ensemble learning, Meta-learning, Clinical feature engineering, Explainable Artificial Intelligence (XAI), SHAP, LIME, Uncertainty quantification
Subjects:	Q Science > QA Mathematics > QA71-90 Instruments and machines
Divisions:	Faculty of Information Science and Technology (FIST)
Depositing User:	Ms Suzilawati Abu Samah
Date Deposited:	28 Oct 2026 03:25
Last Modified:	28 Oct 2026 03:25
URII:	http://shdl.mmu.edu.my/id/eprint/16926

Downloads

Edit (login required)