Predicting antenatal care dropout using machine learning and describing maternal and neonatal outcomes by dropout...

We aimed to develop and validate machine-learning models to predict antenatal care (ANC) dropout and describe maternal and neonatal outcomes by ANC dropout status in a low-resource setting in Northern Sidama, Ethiopia.

Retrospective cohort study.

Health facilities across four districts in Northern Sidama, Ethiopia.

3855 pregnant women aged 15-49 years who initiated ANC between January 2021 and January 2025 were included, whereas women with incomplete records or those who relocated before delivery were excluded.

The primary outcome was ANC dropout, defined as attending <8 recommended ANC visits. Secondary outcomes included maternal and delivery-related outcomes.

Among 3855 pregnant women, 71.4% (n = 2,753) did not complete the recommended eight ANC visits. The extreme gradient boosting (XGBoost) model achieved the strongest predictive performance, with an accuracy of 87%, area under the receiver operating characteristic curve of 0.91, sensitivity of 85% and specificity of 88%. The model identified late ANC initiation (>16 weeks), rural residence, low maternal education, distance to health facilities (>5 km), multiparity and lack of reliable transportation as the main contributors to predicted ANC dropout. Women who did not complete ANC showed higher proportions of maternal complications, including pre-eclampsia (12%) and antepartum haemorrhage (8%), and were less likely to deliver in health facilities (41% vs. 87%). Neonatal adverse outcomes, such as low birth weight and need for resuscitation, were also more frequent among women who did not complete ANC. All outcome comparisons are descriptive and do not imply causal relationships.

Machine-learning models, particularly XGBoost, effectively predicted ANC dropout and identified women with a high predicted probability of dropout. We also described differences in maternal and neonatal outcomes by ANC dropout status. These findings support the potential use of predictive tools to guide early identification and targeted maternal health interventions in resource-limited settings.