Application of AI to Ultrasonographic Images to Aid the Clinical Care of Pregnant Women With Pre-eclampsia in Uganda: A Protocol for a Pilot Study

TL;DR

This study aims to develop an AI tool using Doppler ultrasound images to help predict pre-eclampsia complications in pregnant women in Uganda.

Contribution

It creates one of the first annotated obstetric Doppler ultrasound datasets from sub-Saharan Africa.

Findings

A dataset of Doppler ultrasonographic images linked to clinical metadata will be generated.

A machine learning model will be developed to predict maternal and fetal complications of pre-eclampsia.

Abstract

Background

Artificial intelligence (AI) refers to computer systems designed to perform tasks requiring human intelligence, including medical diagnosis. AI methods have enhanced diagnostic processes across various diseases. In obstetrics, conditions such as pre-eclampsia are typically assessed using USG, yet access to these services and trained sonographers remains limited. Automated diagnosis using AI applied to stored images offers an opportunity to improve maternal and fetal outcomes. In Uganda, progress in integrating AI into obstetric care has been minimal, despite the high burden of complications. This study aims to create a Doppler USG image database, annotated for machine learning models to predict pre-eclampsia complications.

Methods

This cross-sectional study will enroll 150 pregnant women seeking obstetric USG services at Divine Mercy Hospital, Mbarara City, Uganda. Participants will be recruited consecutively from the outpatient department, with half expected to have pre-eclampsia. Sociodemographic, obstetric, and clinical data will be systematically collected, de-identified, and linked to corresponding Doppler ultrasonographic images. Data elements will include acquisition parameters, key Doppler indices, and anonymized demographic information. All data will be securely stored in a structured repository hosted by the Data Management and Analysis Core (DMAC) of the Mbarara University Data Science Research Hub. Images will be annotated using a standardized protocol by trained experts and linked with structured clinical metadata. The dataset will be partitioned into training, validation, and test subsets. Machine learning approaches will include convolutional neural networks, ensemble learning, and transfer learning. Performance will be evaluated using cross-validation, area under the receiver operating characteristic (ROC) curve (AUC), precision, recall, and F1-score, with hyperparameter optimization via grid and Bayesian search.

Results

The primary outcome will be a de-identified, annotated dataset of obstetric Doppler ultrasonographic images linked to structured clinical metadata, including the amniotic fluid index, fetal heart rate, umbilical and cerebral arteries, uterine arteries, and placenta. Images will undergo standardized preprocessing and dual expert review, with discrepancies adjudicated by a third reviewer. The secondary outcome is a multimodal AI tool predicting maternal and fetal complications of pre-eclampsia. Data collection targets 150 participants, funded by the National Institutes of Health (NIH) Data Science Initiative for Africa.

Discussion

In settings with limited access to expert imaging, this study will develop a high-quality annotated dataset and a context-specific machine learning model trained on images from pre-eclamptic and non-pre-eclamptic pregnancies. It will be among the first obstetric AI datasets derived entirely from an African population, addressing global gaps in representation. Aligned with Uganda’s Digital Health Strategy, the study supports innovative tools to strengthen clinical decision-making, improve maternal and perinatal outcomes, and advance AI integration into routine obstetric care in low-resource settings.

Conclusion

This study will generate one of the first annotated obstetric Doppler ultrasound datasets from sub-Saharan Africa. The findings will inform potential integration of AI tools into routine maternal healthcare.

Figures