An Interpretable Machine Learning Framework for Early Ectopic Pregnancy Prediction Using Routine Clinical Variables

Honi, Dhafer G. and Mohammed, Ahmed Abed and Szathmary, Laszlo (2026) An Interpretable Machine Learning Framework for Early Ectopic Pregnancy Prediction Using Routine Clinical Variables. In: Proceedings of the 13th International Conference on Applied Informatics. Líceum Kiadó, Eger, pp. 146-160. ISBN 9789634963271

Preview

Text ICAI2026-pp146-160.pdf - Published Version Download (1MB) | Preview

Abstract

Ectopic pregnancy (EP) remains a major cause of first-trimester maternal morbidity and mortality, while early diagnosis is challenging because clinical symptoms frequently overlap with those of other gynaecological conditions. This study proposes an interpretable machine-learning framework for early EP risk stratification using routinely available clinical variables. The proposed framework was evaluated on a cohort of 2,060 patients comprising 1,030 EP cases and 1,030 matched controls described by age and 12 binary clinical variables. To ensure unbiased evaluation, the data were divided into independent training and test sets before model development. Synthetic samples were generated exclusively within the training partition using a class-conditional Gaussian copula model to augment model training while preserving the statistical characteristics of the original data. All validation, calibration, statistical analyses, and final performance evaluation were performed exclusively on previously unseen real patient data. Six machine-learning classifiers, including Logistic Regression, Random Forest, Gradient Boosting, Support Vector Machine, Multilayer Perceptron, and a soft-voting ensemble, were evaluated using discrimination, calibration, clinical utility, and explainability analyses. Performance was assessed using ROC-AUC, PR-AUC, calibration metrics, decision-curve analysis, SHAP interpretation, and statistical comparisons based on the DeLong, McNemar, and Friedman tests. The soft-voting ensemble achieved the highest performance on the independent test set, with an ROC-AUC of 0.609 (95% CI: 0.563–0.653). Multivariable logistic regression and SHAP consistently identified prior ectopic pregnancy as the strongest predictor (adjusted OR = 27.8, 95% CI: 8.7– 88.8), followed by psychiatric disease and previous genital surgery. Although predictive performance from routine clinical variables alone was modest, the proposed framework provides transparent, reproducible, and well-calibrated risk estimation. The findings further demonstrate that synthetic data are effective for training augmentation while preserving the independence and validity of real-world clinical evaluation.

Actions (login required)

View Item