AI model accurately detects maternal health risks with great precision

The study argues for integrating AI-driven models into maternal health systems, especially in resource-constrained settings. By embedding risk classification models in mobile health apps or electronic health record (EHR) systems, midwives and clinicians can receive real-time alerts and recommendations based on the physiological data collected during routine visits.

Machine learning models can dramatically improve the classification and early detection of maternal health risks using simple physiological data, providing an accessible and highly accurate tool to support midwifery and prenatal care, according to a new study published in Healthcare.

The findings of the study titled “Maternal Health Risk Detection: Advancing Midwifery with Artificial Intelligence”, show that Random Forest algorithms significantly outperform other classification models in predicting maternal health risks categorized as low, mid, or high, achieving up to 88.03% accuracy.

Can machine learning accurately detect maternal health risks?

Maternal mortality continues to be a major global challenge, particularly in low- and middle-income countries. With 287,000 maternal deaths reported globally in 2020, early detection remains a top priority. The study explores the use of machine learning (ML) to classify maternal risk based on seven physiological indicators: Age, Systolic and Diastolic Blood Pressure, Blood Sugar, Body Temperature, and Heart Rate. Researchers trained six ML models on a dataset of 1,014 anonymized cases collected through IoT-based risk monitoring systems in hospitals and clinics.

Random Forests, a type of ensemble learning algorithm, emerged as the best-performing model. When tested using a 10-fold cross-validation approach, it yielded 88.03% accuracy, along with 88% True Positive Rate and 88.10% Precision. This outperformed other tested models including Support Vector Machines (75.74%), Decision Trees (78.60%), Fully Connected Neural Networks (68.93%), Multilayer Perceptron (69.03%), and Naïve Bayes (59.07%).

The mid-risk category was found to be the most difficult to classify accurately, with significant overlap in physiological features between low- and high-risk categories. To address this class imbalance, researchers employed SMOTE (Synthetic Minority Oversampling Technique), which improved classification performance by generating synthetic examples of underrepresented classes. The technique notably enhanced the model's ability to identify mid-risk cases, increasing overall prediction accuracy and reducing misclassification.

What factors contribute most to risk classification?

Feature selection analysis identified Systolic Blood Pressure and Blood Sugar as the most significant predictors of maternal health risk. Scatter plots confirmed that high-risk cases generally presented with elevated Blood Sugar and higher Systolic BP, while low-risk cases clustered within normal ranges. Mid-risk cases, however, spanned across both ends of the spectrum, reinforcing the classification challenge posed by physiological overlap.

A statistical analysis using multinomial logistic regression further validated the importance of variables like Diastolic Blood Pressure and Blood Sugar in predicting maternal health risk. It revealed that some variables, such as Age in the 10–14 range, were not statistically significant, while DiastolicBP at 70 and BS at 6.6 showed strong associations with high-risk classifications.

Random Forest models also demonstrated excellent performance in confusion matrix analysis, with high accuracy across all three risk levels. The model correctly classified 95.77% of high-risk instances, 81.77% of low-risk cases, and 83.04% of mid-risk cases after optimization. Compared to other published models, including Gradient Boosted Trees (86%), LightGBM (84.73%), and SVM with GridSearch (86.13%), the Random Forest classifier used in this study provided a more practical balance of performance and computational efficiency.

What are the clinical and operational implications?

The study argues for integrating AI-driven models into maternal health systems, especially in resource-constrained settings. By embedding risk classification models in mobile health apps or electronic health record (EHR) systems, midwives and clinicians can receive real-time alerts and recommendations based on the physiological data collected during routine visits.

Clinical Decision Support Systems (CDSS) powered by machine learning could allow for more personalized care, reduce preventable complications, and promote early intervention. The study emphasizes that accurate classification of mid-risk cases, often overlooked in binary classification systems, could significantly enhance outcomes by enabling timely risk escalation or de-escalation protocols.

The authors also highlight the need for further validation through larger, multicenter datasets. The current dataset was region-specific, possibly limiting generalizability to other populations. Expanding training datasets with demographic, socio-economic, and contextual health information would improve the robustness of the models across different geographies.

The study also highlights ethical considerations. While machine learning provides significant predictive benefits, deploying AI in healthcare settings raises concerns about interpretability, fairness, and informed decision-making. The study calls for human-in-the-loop approaches where AI augments, rather than replaces, clinical judgment.