ABS 082: HealthHive: A Web-Based Maternal Health Monitoring App for Proactive Infant Mortality Prevention

Jaya Kolluri

Van Wickle (2025) Volume 1, ABS 082

Introduction: Infant mortality remains a significant global health issue, often linked to preventable pregnancy complications such as preeclampsia, gestational diabetes, and preterm birth. Children face the highest risk of dying during their first month, with an estimated average of 43 deaths for every 1,000 live births. Many cases of infant mortality are due to underlying health issues detected too late for effective management. Proactive monitoring and timely detection of abnormal maternal health patterns are essential to reducing risks and the infant mortality rate. Identifying early signs of complications, such as abnormal heart rate and blood pressure, can enable healthcare providers and expectant mothers to take preventive measures and potentially save lives.
By using machine learning models to analyze the vast amount of data generated by smart wearable devices, the project aims to identify abnormal patterns in real time, enabling healthcare professionals to offer proactive care. The long-term goal is to develop a comprehensive monitoring system that alerts pregnant mothers to potential risks, equips doctors with timely information, and sheds light on disparities in infant mortality rates across different ethnic and socio-economic groups. While phase one of the research focused on developing machine learning models for anomaly detection, current research work focuses on developing a web-based maternal health monitoring application called HealthHive, which provides a graphical interactive interface for tracking and anomaly detection. Using technologies like PHP and Python, the app features interactive data visualizations and highlights abnormal data points.
This web application bridges the gap between research and clinical practice by offering an intuitive interface for continuous monitoring and proactive risk management. By making advanced analytical tools accessible, this app offers a proactive and scalable solution to a critical public health problem. Extensions for future work will focus on integrating wearable device data and validating the system with real-world data.

Methods: This study utilized a 10-day simulated dataset representing physiological readings from wearable smart devices, capturing heart rate, systolic, and diastolic blood pressure every minute. Normal values were set within clinically accepted ranges, while 5% of the data was intentionally varied to simulate anomalies. To address the challenges of unlabeled real-time health data, we applied unsupervised machine learning models—specifically Isolation Forest and Local Outlier Factor. These models were incrementally trained using prior days’ data to detect anomalies in new inputs, simulating real-world personalization. The goal was to evaluate which model could better identify abnormal trends that may signal health risks in pregnancy. The methodology was designed to prioritize HIPAA compliance by not sharing data across simulated patients and to test the impact of incremental learning over time. Performance metrics were recorded daily to observe model behavior and accuracy across increasing volumes of training data, noting any signs of overfitting or degradation.

Results: The Isolation Forest model consistently outperformed the Local Outlier Factor in detecting anomalies in simulated maternal health data. Accuracy improved steadily as more data was introduced, peaking by the seventh day. Beyond this point, performance plateaued, with signs of overfitting. This supports our hypothesis that unsupervised incremental learning can personalize anomaly detection in real-time physiological data. These results validated the feasibility of a smart-device-based alert system for expectant mothers and suggest that careful control of data input length can optimize performance while avoiding overfitting. The approach shows promise for scalable, privacy-conscious maternal health monitoring.

Discussion: Our findings highlight the potential of unsupervised, incremental learning models—particularly Isolation Forest—for real-time, personalized maternal health monitoring using smart device data. This approach enables early detection of abnormal physiological patterns without requiring labeled datasets, addressing a major barrier in clinical deployment. The success of anomaly detection up to day seven suggests a practical window for adaptive learning before overfitting occurs. These results open the door for scalable, privacy-preserving monitoring tools that can assist both patients and healthcare providers. Future work will focus on expanding to real-world datasets, integrating additional health parameters, and exploring interventions based on detected anomalies.