Machine learning meets maternal health: Uncovering spatial blind spots in antenatal care quality in Bangladesh.

BACKGROUND: High-quality antenatal care (ANC) is defined as four or more antenatal visits with at least one to a medically trained provider, measurement of weight and blood pressure, testing of blood and urine, and receipt of information on potential danger signs at least once during pregnancy. Though Bangladesh has almost universal ANC coverage, there is widespread inequality in the quality of these services. Traditional statistical models utilized in studies have tended to disregard complicated interconnections between socio-demographic, service-based, and regional factors that influence the quality of ANC. Using nationally representative data, this paper applies machine learning (ML) approaches to classify ANC quality, identify regional hotspots of low-quality care, and its factors. METHODS: This study used data from the 2022 Bangladesh Demographic and Health Survey (BDHS), with a sample of 4587 women aged 15-49 who received ANC services. To predict binary ANC quality outcomes (high vs. low), three models were used: logistic regression, random forest (RF), and gradient boosting machine (GBM). Class imbalance was addressed using the ROSE (Random Over-Sampling Examples) technique, and model performance was evaluated using accuracy, sensitivity, specificity, and area under the ROC curve with 5-fold cross-validation. The most influential predictors were identified using feature importance analysis, and projected probabilities were aggregated at the cluster and division levels for spatial hotspot analysis. Geographic mapping was then utilized to demonstrate regional differences. RESULTS: The GBM model outperformed the others, with the greatest prediction value (accuracy: 81.3%, sensitivity: 70.6%, specificity: 84.7%, AUC-ROC: 0.889). Number of ANC visits, wealth index, place of residence, maternal education, and media access were all significant predictors. Spatial studies found hidden regions with high ANC visit coverage but low predicted ANC quality, highlighting considerable spatial differences in service quality. These hotspots are concentrated in Rangpur and Sylhet, which are far from Dhaka, the capital of Bangladesh, demonstrating spatial disparities in the usage of ANC services. CONCLUSIONS: The study shows that machine learning can classify ANC quality and reveal spatial disparities, aiding policymakers in targeting programs and allocating resources.