Blood Pressure Monitoring and Assisting System to Support Pregnant Women at High-Risk for Preeclampsia

Authors:

Abstract
Preeclampsia is a complex pregnancy condition characterized by high blood pressure, often leading to damage to organs and tissues. Early detection and management of preeclampsia significantly reduce the risk of serious complications including seizures, liver failure and heart disease in both mother and baby. This paper aimed to provide tracking of blood pressure levels in high-risk pregnant women. The solution serves as a valuable tool in supporting pregnant women at high-risk for preeclampsia, enabling early detection and effective management of the condition. Using K-fold cross-validation, the proposed model achieved clinically acceptable prediction accuracy, with systolic blood pressure (SBP) and diastolic blood pressure (DBP) values estimated within an average error margin of approximately 0.414 mmHg (MAE) and a validation loss of $\text{0. 5 4 4}$ (MSE) for ResNet50, compared to $\text{0. 4 1 9 ~ m m H g}$ (MAE) and $\text{0. 5 4 2}$ (MSE) for EfficientNetB0. The results indicate that ResNet50 achieved slightly better prediction accuracy in terms of mean absolute error, suggesting that its deeper residual connections capture more discriminative features, whereas EfficientNetB0 achieved a marginally lower overall validation loss, due to its compound scaling and stronger regularization. Both models, however, yielded negative $\mathrm{R}^{2}$ values (-0.291 for ResNet50 and -0.309 for EfficientNetB0), indicating poor variance explanation and performance below a naive baseline. The novelty of this study resides in its end-to-end integration of a CNN-based blood pressure estimation algorithm with a mobile application specifically built for preeclampsia support. Unlike existing remote blood pressure monitoring techniques that rely on wearable equipment or human input, this system offers non-contact, camera-based assessment with typical smartphones while including clinical context.

This publication has 19 references indexed in Scilit: