An AutoML-based stacked ensemble framework for early prediction of maternal health risks

Abstract Predicting maternal health risks is a major clinical challenge. Early and precise predictions can reduce maternal deaths and complications. AI applications in maternal healthcare play a vital role in predicting risk and improving diagnoses. However, conventional machine learning approaches to maternal risk prediction require extensive manual effort during model design, algorithm selection, training, and tuning, which causes delays and is prone to human error. This study overcomes these challenges by implementing an automated machine learning (AutoML) approach based on the H2O framework in predicting pregnancy risks. The dataset used is a publicly available MHRD consisting of 1014 observations and six clinical variables. In total, 52 different ML models based on six different algorithms were automatically trained and tested in one unified pipeline. These algorithms include XGBoost, random forest, generalized linear model, and stacked ensembles. The stacked ensemble model demonstrated superior predictive performance on all other base models. Also, this model showed an area under the curve value of 0.9539, LogLoss of 0.4093, and root mean square error value of 0.3638 when applied to a holdout test set. The model achieved 96.2% sensitivity for high-risk pregnancies, which is an essential characteristic in clinical practice. These results outperform all individually trained base learners and previously reported results from manually tuned models in the literature. This work demonstrates that the AutoML-based framework for maternal risk prediction is capable of automatically selecting the best machine learning models from clinical data. The framework also enhances predictive accuracy, decreases development effort, and facilitates prompt clinical decision-making in maternal healthcare.