Maternal Health Risk Classification

Overview

Maternal health risk refers to the level in which a mother or child is likely to experience harmful outcomes during the pregnancy or labor period. Despite medical advances, it remains a global issue. This study aimed to improve maternal health risk classification by testing multiple machine learning methods.

School

College of Arts & Sciences

Program

BS in Data Science

Researcher

Lily Vogel '25

Data Science

College of Arts & Sciences

Maternal Health Risk Classification

Background

Maternal health risk refers to the likelihood that a mother or child experiences harmful outcomes during the pregnancy, labor, or postpartum period (Togunwa et al., 2023).

Unfortunately, maternal health risk remains a global concern. According to the World Health Organization, nearly 800 women died every day on average from preventable issues during pregnancy and childbirth in 2020 (WHO, 2024).

Machine learning models can recognize complex patterns within medical data that could be overlooked by humans or traditional statistical analyses (Sathasivam and Abdullahi, 2024).

While previous studies have explored machine learning techniques for maternal health risk classification, there are still limitations in model variation and feature importance analysis.

Introduction

Given gaps in research, this study aimed to improve these issues by examining multiple ML methods to assess maternal health risk using a dataset with more maternal health features to identify the most crucial ones.

In this study, I aimed to improve maternal health risk classification by addressing two research goals:

First, I used three machine learning models – an ordered probit logistic regression, a decision tree, and a random forest – to identify the most important factors for classifying maternal health risk.
Second, I evaluated and compared overall model accuracy as well as within- group performance using confusion matrices.

By understanding key risk factors and improving accuracy, this study contributes to the effort to use machine learning to improve maternal healthcare outcomes.

Methods

Dataset

All data came from the Maternal Health Risk dataset from the University of California, Irvine (UCI) Machine Learning Repository (Ahmed, 2020).
Data was collected from hospitals, community clinics, and other healthcare facilities in Bangladesh.
It contained six features, including age, systolic blood pressure, diastolic blood pressure, blood glucose levels, body temperature and heart rate.
The target variable, risk level during pregnancy, had three classes – low, mid, and high risk.

Predictors of Maternal Health Risk

Fit 3 optimized models – ordered probit logistic regression, decision tree, and random forest.
Found most important predictors for classifying risk in each model.

Model Comparison

Compared the accuracy of each of the 3 models.
accuracy = sum of correct classifications / total classifications
Analyzed within-group classification accuracies by creating confusion matrices.

Results

Graph depicting the heatmap of feature importance rankings, with dark blue equating to lower rank values and dark red equating to higher rank values

Figure 1. Heatmap of feature importance rankings

Lower rank values indicate greater importance (dark blue),
while higher rank values indicate lower importance (dark
red).

• Blood glucose level (BS) consistently ranked as the most
important feature across all models.
• Heart rate had a moderately high or the highest rank in
all three models, indicating its low significance.

Table showcasing the model accuracy comparison based on the overall prediction accuracy for each model

Figure 2. Model accuracy comparison based on the overall
prediction accuracy for each model

The random forest model achieved the highest accuracy at
84.24%, followed by the decision tree model with 64.53%.
The logistic regression model had the lowest accuracy, with
a performance of 58.13%

Graph depicting combined confusion matrix showing the aggregated classification results from all 3 models

Figure 3. Combined confusion matrix showing the
aggregated classification results from all 3 models.

It shows predictive performance for classifying into low,
mid, and high classes. Each row of the matrix represents the
true class, while each column represents the predicted
class. Darker blue tones indicate more instances classified in
that cell. The diagonal cells indicate correct classifications
for each class, while the off-diagonal elements represent
misclassifications.

Overall, the low and high risk classes were classified most
accurately, with 191 and 130 correct predictions,
respectively.

Discussion

Predictors of Maternal Health Risk

Blood glucose level was the most important predictor for classification.
Systolic blood pressure was the second most important predictor.
Heart rate was the least important feature for classification as it seemed to contribute to the models the least overall.

Model Comparison

The results suggest a substantial difference in classification performance between the models, with random forest outperforming the other two models by a considerable margin.
The decision tree model demonstrated moderate performance, ranking second in terms of accuracy.
The ordinal logistic regression model yielded the worst accuracy out of the three models.

Within-Group Classification

Low and high-risk classes were classified most accurately.
Classifying mid risk was difficult for the models – over half were misclassified, with most being wrongly classified as low risk.
Overall, it suggests that the model underestimates moderate risk.

Research Limitations

Since all the data was collected from a single country, the dataset lacked geographic diversity. This suggests that the results may not be generalizable to the whole population.
The dataset lacked social, economic, and psychological variables, which could be important in maternal health risk prediction.

Future Directions

In the future, there are several different ways I could extend my research. Specifically, I would like to:

Improve model performance by experimenting with hybrid models and tuning more hyperparameters.
Test model generalizability by using a dataset from a different region.
Apply the same research goals to a dataset containing multiple demographic features to explore feature importance across different patient groups.
Integrate information from Internet of Things (IoT) devices such as data from wearable devices to develop real-time risk evaluation.

For Further Discussion

This serves as an overview of the project and does not include the complete work. To further discuss this project, please email Lily Vogel.

Course Overview

DS 480: Data Science Capstone serves as a culminating experience for the Data Science major. Students work on an independent project that will allow them to integrate knowledge from their previous courses in the major and apply that knowledge to a problem in a domain of their interest.

Explore Our Areas of Interest

We've sorted each of our undergraduate, graduate and doctoral programs into unique Areas of Interest. Explore these categories to discover which programs and delivery methods best align with your educational and career goals.

Explore Business and Finance at Quinnipiac

Explore Computing and Technology at Quinnipiac

Explore Data and Analytics at Quinnipiac

Explore STEM Programs at Quinnipiac