A data-driven analysis of maternal health risk indicators using machine learning techniques

Introduction

Background

Maternal health is a crucial element of global public health, as it significantly impacts both maternal and infant outcomes. A healthy pregnancy and childbirth are essential not only for the well-being of women but also for the health of future generations (1,2). However, pregnancy remains a period of vulnerability for many women, since complications that arise during this time, such as hypertension, diabetes, or infections, can lead to long-term health issues, or in some cases, result in maternal or neonatal death. According to the World Health Organization (WHO), approximately 295,000 women died due to complications related to pregnancy and childbirth in 2017, with most of these deaths occurring in low-resource settings (3). Although many of these complications are preventable, barriers such as limited access to quality healthcare and delays in medical interventions persist, especially in regions where healthcare infrastructure is inadequate (4,5). Therefore, identifying maternal health risks early has become a vital goal in improving maternal health outcomes (6).

Several clinical indicators play an important role in determining a pregnant woman’s health and the risks of developing complications. For example, elevated systolic and diastolic blood pressure is strongly associated with preeclampsia, a dangerous condition that, if left untreated, could result in severe complications or even maternal death (7,8). Likewise, high blood sugar (BS) levels during pregnancy may indicate gestational diabetes, which can pose risks to both the mother and the baby if not properly managed. By closely monitoring key health indicators such as blood pressure, BS, and heart rate, healthcare providers can intervene early and potentially prevent serious complications from arising (9,10). However, accurately assessing and predicting these risks remains challenging, particularly when dealing with diverse populations where risk factors and access to care vary significantly. As a result, there is an urgent need to explore advanced methods that can enhance the accuracy of maternal health risk assessments (11,12).

This challenge of processing multiple risk factors simultaneously while accounting for population diversity has led to growing interest in advanced analytical approaches, particularly machine learning (ML). In recent years, healthcare has increasingly relied on data-driven approaches to analyze and predict health outcomes. Although traditional statistical methods have long been used to evaluate risk factors, the rise of ML has expanded the possibilities for developing more precise predictive models (13,14). Statistical methods provide a detailed analysis of data distributions, correlations, and variance, while ML techniques enable the discovery of hidden patterns and relationships that might not be visible through standard statistical techniques (15). When combined, these methods can offer powerful predictive models that provide early warnings about maternal health risks. These models also enable healthcare providers to deliver more personalized interventions, thus contributing to more equitable and effective maternal care (8,16). ML has shown great promise in healthcare, particularly for tasks like risk prediction and classification. Algorithms such as logistic regression, support vector machines (SVMs), and Decision Trees are frequently employed in healthcare settings to predict disease outcomes and classify patients into different risk categories (17,18). These algorithms are especially useful for identifying complex patterns in health data that may be too subtle for traditional methods to detect (19). However, the accuracy and reliability of these classification models can depend heavily on the quality and quantity of data used for training. Poor data quality or insufficient data can lead to decreased model performance and may limit the generalizability of the findings (20). For instance, in maternal health, ML can be used to classify women into various risk levels based on physiological indicators like blood pressure, heart rate, and BS levels (21,22).

The primary objective of this study is to provide an educational demonstration of how ML techniques can be applied to maternal health data for exploratory analysis. Specifically, this research seeks to explore the relationships between key physiological factors, such as systolic and diastolic blood pressure, BS, body temperature, and heart rate, and maternal health risk levels. Logistic regression will be applied as a feature selection method to identify and retain the most significant factors that influence maternal risk, which will then be used to improve the predictive performance of the ML models. It is important to note that this study does not attempt to make definitive clinical conclusions but rather focuses on illustrating the utility of ML techniques in a maternal health context. Notably, limited studies have integrated logistic regression for feature selection followed by the application of ML algorithms in the context of predicting maternal health risks. This exploration aims to provide insights into how data-driven techniques can improve early risk detection and intervention strategies in maternal healthcare, with the understanding that further validation by clinical experts and the use of more comprehensive datasets would be necessary for real-world applications.

Literature review

Maternal health risk indicators are essential for understanding and improving health outcomes for mothers, as they help identify potential risks and enable timely interventions to prevent maternal morbidity and mortality (23). However, in both medical and public health research, key indicators include the maternal mortality ratio (MMR), maternal near-miss cases, as well as various socio-demographic and clinical factors. The MMR, a widely used measure, tracks the number of maternal deaths per 100,000 live births, making it a crucial indicator for comparing maternal healthcare quality across countries. Globally, there has been a significant decline in MMR, with a 38% reduction recorded since 2017. Furthermore, the global target for 2030 is to ensure that no country has an MMR higher than 140 per 100,000 live births (24). Several factors influence MMR, including maternal age, education level, and pregnancy-related medical conditions. For instance, women over the age of 35, those with limited education, and those facing complications such as obstructed labor or hypertensive disorders are at a higher risk of maternal mortality (25). In addition to MMR, maternal near miss (MNM) cases provide valuable insights into maternal health, particularly in regions where maternal mortality has decreased. MNM refers to situations where women experience severe complications during pregnancy or childbirth yet survive. This indicator is becoming increasingly important in maternal health research, as it reflects the ability of healthcare systems to manage life-threatening conditions (26,27). Significant predictors of MNM include maternal age, education level, antenatal care follow-ups, and the presence of medical conditions during pregnancy (26). Moreover, the concept of lifetime risk of MNM offers a way to estimate a woman’s cumulative risk of experiencing severe maternal morbidity throughout her reproductive years (27).

Socio-demographic and clinical factors also play a critical role in shaping maternal health outcomes. Many studies have analyzed the relationship between Socio-demographic and clinical factors and maternal health. These included, in their study showed that older maternal age and lower levels of education are associated with higher risks of adverse outcomes (25). Moreover, Akor et al. [2024] conducted a study in Zamfara State, Nigeria, and found that higher levels of education and income are closely linked to better access to maternal healthcare services. Their research demonstrated that women with greater educational attainment and higher income levels were more willing and able to access maternal healthcare (28). This finding highlights the significant role that socio-economic status plays in shaping maternal health outcomes. Minopoli et al. [2024] explored the impact of ethnicity and socioeconomic deprivation on pregnancy outcomes and revealed that these factors are critical determinants of adverse maternal health outcomes. The results revealed that Black and South Asian women are at higher risk for adverse outcomes compared to White women. Additionally, it was noted that socioeconomic deprivation further exacerbates these risks (29). Similarly, Shen et al. [2022] and Alfred and Wilson [2022] examined racial and ethnic disparities in the U.S., where Black women continue to experience disproportionately higher rates of adverse maternal health outcomes. These disparities were attributed, in part, to systemic inequities in quality (30,31). Furthermore, the influence of cultural beliefs and practices on maternal morbidity in Lagos State, Nigeria, has been investigated, revealing that cultural factors, such as traditional practices and dietary habits, significantly impact maternal health, particularly in conditions like high blood pressure and diabetes during pregnancy (32).

Many analytical tools have been employed to assess maternal health risks. Traditional methods for maternal health risk analysis have largely depended on clinical assessments and statistical models. Mutlu et al. [2023] and Raihen and Akter [2024] explained that clinical assessments involve healthcare providers evaluating risk factors based on medical history, physical examinations, and laboratory tests. In their approach, parameters such as maternal age, blood pressure, heart rate, and pre-existing conditions like diabetes or hypertension are commonly assessed (33,34). Additionally, Thakkar et al. [2024] and Mustamin et al. [2023] explored the use of basic statistical models for predicting maternal health risks. These models typically involve analyzing historical data through regression analysis and other statistical techniques to estimate the likelihood of complications (35,36). However, while these methods have provided insights, they have notable limitations. First, it has been shown that traditional methods often suffer from limited accuracy, as clinical assessments can be subjective and may fail to account for the complex interactions between various risk factors, which in turn reduces the effectiveness of risk prediction (36). Moreover, traditional statistical models struggle to handle large datasets with multiple variables, which can lead to oversimplified conclusions and result in the misclassification of risk levels. Therefore, some high-risk pregnancies may not receive the adequate care they require due to these misjudgments (34,37). Lastly but not least, traditional methods are often criticized for being resource-intensive, as they require significant time and expertise from healthcare professionals. This issue becomes particularly problematic in resource-limited settings, where access to skilled personnel and advanced diagnostic tools is scarce, further complicating the provision of maternal healthcare (35).

ML models, including logistic regression, Decision Trees, K-nearest neighbors (KNN), and SVM, have demonstrated significant potential in enhancing the prediction and early detection of maternal health risks. Logistic regression is widely used due to its simplicity and interpretability in predicting binary outcomes, offering specific advantages in maternal healthcare settings. Its primary strength lies in providing clear interpretability of results, allowing healthcare providers to understand how specific factors like maternal age, blood pressure, or previous complications influence risk predictions. However, its performance is often limited by its assumption of linear relationships between variables, which may not capture the complex interactions present in maternal health data. Studies have shown that while logistic regression provides a good baseline for risk prediction, its performance is often surpassed by more complex models that can capture intricate data patterns (37,38). Decision Trees offer valuable approach for maternal health risk assessment due to their alignment with clinical decision-making processes. Their hierarchical structure mirrors how healthcare providers naturally approach patient assessment, making them especially useful for risk stratification in maternal care. Research has shown that Decision Trees are effective in classifying maternal health risks into categories like high, medium, and low, providing an intuitive tool for healthcare providers. However, they can be prone to overfitting, especially with limited data, and may create overly complex trees that don’t generalize well to new patient populations (38,39). KNN ML algorithm provides a non-parametric classification method that classifies data points based on the majority class of their nearest neighbors (40). Although researchers have used KNN in maternal health risk prediction, it generally performs worse compared to models like random forests and SVMs (37). On the other hand, SVMs are known for their ability to handle high-dimensional data and are particularly useful when dealing with data that is not linearly separable (41). A study conducted by Raihen and Akter [2024] demonstrated that SVMs can achieve high levels of accuracy, especially when optimization techniques such as cross-validation are employed to fine-tune model parameters (34).

Ensemble methods, including random forests and gradient boosting, have consistently outperformed individual ML models in maternal health risk prediction. These methods improve accuracy and robustness by combining the strengths of multiple models. To illustrate, Random forests have been shown to effectively capture complex data patterns, leading to high predictive accuracy in maternal health risk scenarios (36,39). In addition, advanced techniques such as principal component analysis (PCA) for feature extraction, coupled with ensemble voting classifiers, have further enhanced model performance. One study utilizing PCA, and a stacked ensemble model reported an impressive accuracy of 98.25%, underscoring the potential of combining ML with feature engineering to achieve superior predictive outcomes (42).

Pawar et al. [2022] developed a robust ML model to predict maternal health risks, emphasizing the importance of maternal health during pregnancy, childbirth, and the postpartum period. The study aimed to tackle high maternal morbidity and mortality rates by leveraging ML for early prediction and intervention. Using a dataset from the UCI Machine Learning Repository, which included features such as age, blood pressure, BS, body temperature, and heart rate, the authors identified these indicators as essential for risk assessment. Their model achieved an accuracy of 70.21% across worst, average, and best-case scenarios by applying feature selection with the Gini index and using k-fold cross-validation for robustness. Performance metrics, including accuracy, precision, recall, and ROC, demonstrated the model’s ability to handle noisy and adversarial data. Highlighting key causes of maternal mortality, such as hemorrhage and high blood pressure, the study stressed the need for effective predictive strategies. Pawar et al. concluded that ML can significantly enhance maternal health outcomes (43).

Methods

Data overview and preprocessing

In this section, the dataset description and preprocessing steps undertaken in this study are presented. The analysis followed a structured process, starting with data preprocessing and culminating in the deployment of ML models. Figure 1 illustrates the analysis flowchart of this study.

Figure 1 Flowchart for the prediction of maternal health risk.

Dataset description

The dataset utilized in this study was sourced from the UCI Machine Learning Repository and focuses on maternal health risk indicators. It comprises anonymized patient records annotated with relevant features for predicting maternal health outcomes. The dataset includes physiological features such as age, blood pressure, blood glucose levels, body temperature, and heart rate, as detailed in Table S1. In this dataset, the outcome variable ‘risk level’ categorizes maternal health risk into three levels: low risk, mid risk, and high risk. This classification is based on key physiological indicators such as blood pressure (systolic and diastolic), BS levels, body temperature, and heart rate. Each patient record in the dataset was assigned a risk level based on predefined thresholds or clinical criteria, though the dataset itself does not explicitly document the exact methodology used for labeling. Given this limitation, our study treats ‘risk level’ as a general indicator of overall maternal health risk rather than a specific clinical diagnosis.

Data loading and cleaning

The initial step involved importing the dataset into the Jupyter Notebook analysis environment and checking for completeness. This included addressing missing values and identifying outliers.

Handling missing values and outliers

Missing values, if present, were treated either via imputation or removal. Imputation involves replacing missing values with statistically derived substitutes, such as the mean, median, or mode of the corresponding feature, to preserve the integrity of the dataset without discarding valuable data (44). For outliers, detection was performed using the interquartile range (IQR) method, a statistical technique that identifies data points significantly deviating from the central 50% of the distribution. Specifically, the IQR is calculated as the difference between the 75th percentile (Q3) and the 25th percentile (Q1) of the data. Data points lying outside 1.5 times the IQR above Q3 or below Q1 are considered outliers and were either removed or treated to reduce their influence on model performance (44).

Encoding categorical variables

Categorical variables were minimal in this dataset, with the ‘risk level’ being the only categorical attribute. It was processed using label encoding, a method that converts categorical data into numerical format to ensure compatibility with ML algorithms. In label encoding, each unique category is assigned a distinct integer value. For instance, in this dataset, the ‘Risk Level’ categories (e.g., “low risk”, “mid risk”, and “high risk”) were encoded as 0, 1, and 2, respectively. This transformation allows algorithms to interpret categorical data as numerical input without altering the inherent relationships between categories (45).

Standardizing continuous variables

Continuous variables, including attributes such as diastolic blood pressure, body temperature, and heart rate, were standardized using Z-score normalization. This method involves centering each feature around its mean and scaling it to unit variance, ensuring that the features are on a comparable scale (46). The first five rows of the preprocessed dataset after standardization and label encoding are shown in Table S2.

Feature selection

Feature selection was conducted using a backward elimination process in logistic regression. This method systematically removes predictors from the model to identify the most statistically significant features. Predictors with P values greater than 0.05, indicating low statistical significance, were sequentially excluded, leaving only the most relevant features for predicting maternal health risk. Backward elimination was chosen for its effectiveness in refining models by reducing complexity and improving interpretability. By focusing on statistically significant predictors, the method minimizes the risk of overfitting, enhances computational efficiency, and ensures that the final model includes only features with a meaningful contribution to the target variable (47).

Data splitting

Finally, the dataset was divided into training (75%) and testing (25%) sets to evaluate model performance and ensure generalizability on unseen data (48).

ML techniques

This section highlights the use of various classifiers in the study, along with the implementation of hyperparameter tuning for each model to enhance their performance.

Decision Trees algorithm

In this research, Decision Trees were utilized to build an interpretable model for predicting maternal health risks, leveraging their capability to handle numerical and categorical data. The model constructs a tree-like structure, where each decision node represents a feature split, such as age, blood pressure, or glucose level, to arrive at an outcome. This approach helps identify critical risk factors by evaluating the features that contribute the most to information gain at each node. Hyperparameter tuning was conducted using GridSearchCV, which facilitated a comprehensive search over predefined parameter grids, including criterion (either ‘gini’ or ‘entropy’), max_depth, min_samples_split, and min_samples_leaf. The process utilized 5-fold cross-validation to evaluate different combinations of these parameters, aiming to identify the optimal configuration for the model. The identified best hyperparameters were then applied to refine the final Decision Tree, improving its accuracy and overall predictive performance. The evaluation process included calculating metrics like accuracy, precision, recall, and F1-score to assess model efficacy.

Logistic regression algorithm

Logistic regression was applied due to its capability to handle multi-class classification problems, such as predicting maternal health risk levels, which are categorized into multiple risk levels rather than a binary outcome. The model uses a logistic function to estimate the probability of a patient falling into each risk category based on several clinical features. In addition to serving as a predictive model, logistic regression was used for feature selection. By examining the statistical significance of each variable’s coefficient, we identified and retained only the most relevant predictors of maternal health risk. This feature selection process reduces model complexity and improves interpretability by focusing on key variables. Logistic regression is particularly valuable in medical applications as it provides interpretable coefficients that help quantify the contribution of each variable, giving healthcare professionals insight into how specific factors influence the predicted risk level.

To enhance the model’s performance, hyperparameter tuning was performed using GridSearchCV. The tuning process explored a grid of values for the regularization strength parameter “C”, along with different solvers (liblinear and lbfgs), and various values for maximum iterations. This comprehensive search identified the best combination of parameters that balanced model complexity and accuracy, improving the predictive performance of the logistic regression model for multi-class classification tasks. The probability that a patient falls into a specific maternal health risk level, p(x), can be calculated using Eq. [1] (49).

p(x)=11+e−(β0+β1x1+β2x2+...+βnxn)[1]

Here, β0,β1,...,βn are the coefficients corresponding to each feature x1,x2,...,xn, where n represents the total number of attributes.

KNN algorithm

The KNN algorithm was employed as a non-parametric method to predict maternal health risk levels by comparing new cases to similar instances within the dataset. Utilizing important clinical variables such as age, systolic and diastolic blood pressure, body temperature, glucose levels, and heart rate, KNN identifies the closest neighbors to determine the most frequent maternal risk outcome among them. This algorithm is effective at identifying complex patterns without requiring a predefined model structure, making it well-suited for situations where decision boundaries may be irregular. The prediction is based on the most common maternal health risk level among the nearest neighbors, with distances calculated using multiple metrics such as Euclidean, Manhattan, or Minkowski distances across the selected health attributes. The function for the KNN model is defined as in Eq. [2] (50).

y=mode{yi|i∈k-nearestneighbors}[2]

Here, y represents the predicted maternal health risk level. The distance between cases is typically calculated using the Euclidean distance, as shown in Eq. [3] (50).

d(xi,x)=(xi1−x1)2+(xi2−x2)2+...+(xin−xn)2[3]

Here, xi andx represent the multi-dimensional feature vectors of existing cases in the dataset and the new patient case, respectively. Each dimension corresponds to one of the health variables used to assess heart disease risk.

Hyperparameter tuning for the KNN model involved an exhaustive grid search aimed at determining the optimal number of neighbors, distance metric, and the best weighting function (uniform or distance-based). This comprehensive search was performed to enhance the model’s accuracy by evaluating different combinations of hyperparameters. The process ensured that the final model configuration was optimized for the specific characteristics of the maternal health risk dataset, improving its predictive performance.

SVMs algorithm

SVM was employed to effectively distinguish between different maternal health risk levels. SVM functions by identifying the optimal hyperplane within a high-dimensional space, where each dimension corresponds to a relevant health feature. This method is particularly effective in preventing overfitting, making it well-suited for handling complex medical datasets. Additionally, SVM enhances predictive performance by maximizing the margin between the closest points of each class, improving classification accuracy. The SVM model is mathematically represented in Eq. [4] (51).

f(x)=sign(∑i=1nαyiiK(xi,x)+b)[4]

In this context, αi represents the Lagrange multipliers, yi denotes the class labels, and (xi,x) refers to the kernel function, which calculates the dot product in the transformed feature space. The term “b” signifies the bias. This approach enables SVM to manage non-linear relationships by utilizing various kernel functions to project the input data into higher-dimensional spaces, where linear separation is achievable. The SVM model’s hyperparameter tuning was conducted using a grid search approach, focusing on the regularization parameter “C” and the kernel coefficient “gamma”. Additionally, this method explored various kernel types, including linear, rbf, and poly. The grid search aimed to identify the optimal combination of these parameters that would maximize the SVM’s performance. By evaluating multiple configurations, the process ensured that the final model setup was the most effective for the given dataset.

Evaluation metrics

To evaluate the performance of the models, key metrics such as precision, recall, F1-score, and accuracy were used. These metrics provided valuable insights into how effectively each model predicted maternal health risk levels, helping to determine the most suitable model or combination of models for practical clinical application. The evaluation process involved comparing the predicted results with the actual outcomes from the test set, ensuring the models’ ability to generalize beyond the training data.

Precision (P) measures the percentage of positive predictions that were correct, calculated using Eq. [5] (52). Recall (R) assesses the percentage of actual positive cases that were correctly identified, as described in Eq. [6] (52). The F1-Score, which is the weighted harmonic mean of precision and recall, offers a balanced metric between these two aspects, as outlined in Eq. [7] (52). Finally, Accuracy represents the proportion of overall correct predictions, calculated by Eq. [8] (52).

P=TPTP+FP[5]

R=TPTP+FN[6]

F1=2×P×RP+R[7]

Accuracy=TP+TNTP+TN+FP+FN[8]

Statistical analysis

All statistical analyses were conducted using Python 3.8. Descriptive statistics were computed for key maternal health indicators, including mean, standard deviation, and IQR. The Kolmogorov-Smirnov test was applied to assess the normality of continuous variables. For the feature selection process, multinomial logistic regression with backward elimination was employed to identify statistically significant predictors. Variables with P values less than 0.05 were considered statistically significant and retained in the final model. The significance level (α) was set at 0.05 for all statistical tests. The significance of model coefficients was determined using Wald’s test, and the 95% confidence intervals were reported for all retained predictors.

Results

In this section, we present the results of the data-driven analysis, highlighting the key maternal health risk indicators identified through the application of ML techniques and their correlation with adverse health outcomes.

Descriptive statistics of key maternal health variables

The descriptive statistics for the key maternal health indicators are presented in Table 1. These indicators include age, systolic blood pressure, diastolic blood pressure, BS, body temperature, and heart rate. The dataset contains a total of 1,014 records for each variable. The mean age of the participants is 29.87 years, with a range spanning from 10 to 70 years, which includes ages outside the typical reproductive range. However, this broad age span may reflect limitations in the dataset and should be interpreted with caution when considering maternal health implications. Additionally, the standard deviation of 13.47 suggests considerable variability in the ages of the individuals analyzed.

For the blood pressure indicators, the mean systolic blood pressure is 113.20 mmHg, while the mean diastolic blood pressure is 76.46 mmHg. Both values are within the normal range; however, the standard deviations of 18.40 and 13.89 for systolic and diastolic blood pressure, respectively, reflect a significant spread in the data. The systolic blood pressure ranges from a minimum of 70 mmHg to a maximum of 160 mmHg, and diastolic blood pressure ranges from 49 to 100 mmHg, indicating the presence of both hypotensive and hypertensive cases. Moreover, the mean BS level is 8.73 mmol/L, with a standard deviation of 3.29 mmol/L, and ranges between 6.0 and 19.0 mmol/L. These values suggest a concentration of the data around the mean, with relatively fewer extreme cases. In terms of body temperature, the participants show minimal variation, with a mean of 98.67 °F and a standard deviation of 1.37 °F, reflecting a largely homogeneous distribution, as temperatures range from 98 to 103 °F. Lastly, heart rate values show a mean of 74.30 beats per minute (bpm) and a standard deviation of 8.09 bpm, with a wide range from 7 to 90 bpm. The wide range of heart rate values may reflect the varied cardiovascular health of the population sampled. Altogether, these descriptive statistics provide a comprehensive overview of the dataset, thus laying the groundwork for further analysis of maternal health risk indicators using ML techniques.

The maternal health cases are categorized into three distinct risk levels: “low risk”, “moderate risk”, and “high risk”, representing the distribution of cases based on these classifications. Among the 1,014 data points, the largest proportion of cases falls within the “low risk” category, accounting for over 400 instances. The “mid risk” category follows with approximately 330 cases, while the “high risk” category represents the smallest group with around 280 cases. This distribution suggests that most of the maternal health cases in this dataset are categorized as low to mid risk, with a smaller but significant portion classified as high risk.

Preprocessing the data: label encoding and standardization

Label encoding transformed the categorical variable “risk level” into numerical values: 0 for low risk, 1 for mid risk, and 2 for high risk. Continuous variables were standardized, resulting in all features having a mean of 0 and a standard deviation of 1. For example, the first row in the dataset represents an individual with a standardized age of −0.361738, a systolic blood pressure of 0.913396, and a risk level of 0, indicating low risk. In contrast, the fifth row shows another individual with a risk level of 1, despite a similar age, due to differences in other health indicators such as systolic blood pressure and BS levels.

Feature selection using multinomial logistic regression

Multinomial logistic regression was used to identify the most significant features for predicting maternal health risk levels. Backward elimination was employed, removing variables with P values greater than 0.05, leading to the exclusion of systolic blood pressure and age. The final selected features included diastolic blood pressure, BS, body temperature, and heart rate. These variables were determined to have the strongest association with maternal health risk levels, as indicated by their statistical significance. The coefficients from the multinomial logistic regression analysis are presented in Table 2, which highlights the impact of these features on each risk level. For risk level =1 (mid risk), the negative coefficients for diastolic blood pressure (−1.0628), BS (−2.1727), body temperature (−0.9992), and heart rate (−0.4726) indicate that lower values of these features are associated with an increased likelihood of being classified as mid risk compared to low risk. For example, a one-unit decrease in diastolic blood pressure significantly increases the odds of being classified as mid risk, as evidenced by its large negative coefficient and statistical significance (P<0.05). For risk level =2 (high risk), the coefficients suggest a similar pattern, where decreases in diastolic blood pressure (−0.8542), BS (−1.2466), body temperature (−0.5768), and heart rate (−0.3106) are associated with a higher likelihood of being classified as high risk. The smaller magnitude of the coefficients for risk level =2 compared to risk level =1 suggests that these features have a slightly less pronounced impact when distinguishing between high risk and low risk.

The constant values for both risk levels =1 and 2 are also statistically significant, indicating the baseline probability of being in these risk categories in the absence of other predictors. Overall, these results demonstrate that lower physiological measurements in the selected features are strongly predictive of higher maternal health risk levels.

ML techniques

The dataset was split into training and testing sets using an 80/20 ratio. This ensures that 80% of the data is used to train the model, while 20% is reserved for testing. In addition, a random state of 42 was applied to maintain consistency and reproducibility in the results.

Decision Tree model

The Decision Tree model was optimized using Grid Search Cross-Validation (GridSearchCV) to identify the best set of hyperparameters (53). The parameter grid included variations for the splitting criterion (‘gini’ or ‘entropy’), maximum tree depth, minimum samples required to split a node, and minimum samples required for a leaf node. The grid search was performed across five cross-validation folds to ensure robust model performance and consistent hyperparameter selection.

Following the grid search, the model was trained with the optimal hyperparameters and evaluated on both the training and test datasets. The classification report for the training set showed an accuracy of 90%, while the test set achieved an accuracy of 76.85%. Precision, recall, and F1-scores were reported for each class (0, 1, and 2) across both datasets in Table 3. The decrease in accuracy from the training set to the test set suggests that while the model fits the training data well, it demonstrates some generalization error on unseen data.

Logistic regression model

The logistic regression model was also optimized using Grid Search Cross-Validation (GridSearchCV) to identify the best set of hyperparameters for the classification task. Moreover, the parameter grid included variations for the regularization strength (C), solver type (liblinear or lbfgs), and the maximum number of iterations (max_iter). Grid search was conducted with five cross-validation folds to ensure the robustness of the model selection process. After determining the optimal hyperparameters, the model was trained on the training data and evaluated on both the training and test datasets. The best parameters for logistic regression were found to be C: 1, solver: ‘lbfgs’, and max_iter: 100. The performance of the model on the training data yielded an accuracy of 61%. The classification report on the test data showed an accuracy of 59%, indicating that the model had a moderate generalization capability. These results are presented in Table 4.

KNN

The KNN model, like the previous models, was optimized using Grid Search Cross-Validation (GridSearchCV) to identify the most suitable hyperparameters for the classification task. The parameter grid included variations in the number of neighbors (n_neighbors), the weighting function (uniform or distance), and the distance metric (euclidean, manhattan, or minkowski). In addition, five-fold cross-validation was used to ensure the selected parameters’ reliability and minimize the risk of overfitting. The optimal combination was three neighbors with the Euclidean distance metric and a distance-based weighting scheme. These hyperparameters were then used to train the KNN model, and its performance was evaluated on both the training and test datasets. The classification report on the training data as shown in Table 5 showed high performance, with an accuracy of 88%, and strong precision, recall, and F1-scores across all classes. However, on the test data, the model achieved an accuracy of 74.38%, showing a slight decrease in performance, though it maintained reasonably good results across classes.

SVM

The SVM model underwent optimization using Grid Search Cross-Validation (GridSearchCV), focusing on tuning hyperparameters such as C, gamma, and kernel type (including linear, rbf, and poly). This method, similar to the approaches used in previous models, aimed to find the most effective combination of parameters by evaluating each set over five cross-validation folds. This approach helps ensure the selection of the best-performing model while minimizing the risk of overfitting. The grid search determined that the optimal configuration for the SVM model involved an RBF kernel with C =100 and gamma =1. The model was subsequently trained and evaluated on the training dataset, yielding an accuracy of 80%. Performance metrics, such as precision, recall, and F1-score, were strong across all classes, indicating good classification abilities. When applied to the test set, the model achieved an accuracy of 66.5%, with slightly lower but still balanced precision and recall values. Although the model performed well on the training set, there was a noticeable drop in performance when evaluated on the test data, reflecting a decrease in generalization. Nevertheless, the SVM model maintained balanced class performance, as evidenced by the macro and weighted averages across the different metrics. Table 6 shows the classification performance report for the SVM model.

Comparative analysis of ML model effectiveness

The performance comparison of ML models provides valuable insights into the classification task, highlighting the strengths and limitations of each algorithm. This study evaluated four models: Decision Tree, KNN, SVM, and logistic regression. To assess their effectiveness, key metrics including accuracy, precision, recall, and F1-score were employed. These metrics provide a comprehensive evaluation of the models’ predictive capabilities and their ability to handle the challenges posed by imbalanced data. Table 7 summarizes the complete comparison of the performance metrics for all models.

Among the evaluated models, the Decision Tree demonstrated the highest performance across all metrics, achieving an accuracy of 0.77, a precision of 0.78, a recall of 0.77, and an F1-score of 0.78. This strong performance underscores the Decision Tree’s ability to capture complex relationships within the data while maintaining interpretability. Hyperparameter tuning further enhanced its robustness, reducing the risk of overfitting, a common challenge with this algorithm. These results suggest that the Decision Tree is highly effective for classifying maternal health risks. The KNN model also performed well, achieving an accuracy of 0.74, a precision of 0.76, a recall of 0.76, and an F1-score of 0.76. Its distance-based classification mechanism allowed it to identify local patterns in the data, resulting in balanced outcomes across all metrics. Although KNN slightly underperformed compared to the Decision Tree, its simplicity and ability to capture nuanced relationships make it a strong alternative. The sensitivity of KNN to hyperparameters, such as the number of neighbors and the distance metric, emphasizes the importance of careful parameter selection for optimal performance.

The SVM model, using an RBF kernel, achieved an accuracy of 0.67, a precision of 0.68, a recall of 0.68, and an F1-score of 0.68. While SVM demonstrated robustness in handling non-linear decision boundaries, its performance lagged behind the Decision Tree and KNN models. This result highlights the potential need for further fine-tuning of hyperparameters or experimentation with alternative kernel functions to enhance its classification capabilities. Despite these limitations, the SVM model remains valuable for datasets with high-dimensional features or complex decision boundaries. Logistic regression, as the baseline model, exhibited the lowest performance, with an accuracy of 0.59, a precision of 0.63, a recall of 0.60, and an F1-score of 0.61. As a linear model, logistic regression struggled to capture the non-linear interactions present in the data, which limited its ability to perform well in this classification task. However, it provided a valuable baseline for comparison, offering simplicity and interpretability. Logistic regression remains a reliable option in scenarios where transparency is prioritized over the ability to model complex relationships.

The overall comparison highlights the Decision Tree as the most effective model for predicting maternal health risks, demonstrating superior performance across all metrics. While KNN offered comparable results, particularly in terms of recall and F1-score, it required careful hyperparameter optimization. The SVM model performed moderately well, though it has potential for improvement through parameter tuning. Logistic regression, while less effective in this context, provided a useful baseline for evaluating the relative performance of the other models. These findings suggest that future research could explore ensemble approaches to combine the strengths of these models, potentially achieving improved accuracy, precision, and recall for maternal health risk prediction.

Discussion

This study focused on exploring the application of applying and comparing the performance of four ML models: Decision Tree, KNN, SVM, and logistic regression in predicting maternal health risks. Using key evaluation metrics such as precision, recall, accuracy, and F1-score, the objective was to demonstrate the strengths and limitations of each model when applied to maternal health data. The results from this comparison highlight the importance of selecting the appropriate model based on the complexity of the data and the need for a balanced approach to various performance metrics. The Decision Tree model emerged as the strongest performer, achieving the highest precision, recall, and F1-scores. Its ability to handle complex, non-linear relationships within the data contributed to its high accuracy, making it a robust option for demonstrating complex model behaviors in healthcare applications. Additionally, the interpretability of Decision Trees makes them highly applicable in healthcare settings, as they provide clear insights into which factors are most influential in determining risk. Similarly, the KNN model demonstrated robust performance, particularly in terms of precision and accuracy, although it did not surpass the Decision Tree. KNN’s strength lies in its simplicity and its ability to classify based on the proximity of data points, which is beneficial when relationships between risk factors are local in nature. However, the model’s sensitivity to hyperparameters such as the number of neighbors and distance metrics requires careful tuning to achieve optimal performance (54).

The SVM model, although effective in identifying non-linear patterns in the data, performed moderately compared to the Decision Tree and KNN models. While the RBF kernel allowed the SVM to capture more complex relationships, its overall accuracy and F1-scores were lower than expected. This suggests that further refinement through hyperparameter optimization could enhance its predictive capabilities in maternal health risk prediction. Nevertheless, the SVM model remains valuable in scenarios where separating distinct risk categories with clear boundaries is essential (55). In contrast, logistic regression produced the weakest results in terms of accuracy, precision, and recall. As a linear model, it struggled to capture the more complex, non-linear interactions inherent in maternal health risk factors. Despite its lower performance, logistic regression still has value in scenarios where model simplicity, speed, and interpretability are prioritized. It provides a solid baseline for comparing more sophisticated models and is particularly useful in situations where the relationships between variables are simpler (56).

While this study explores maternal health risk prediction using ML, it is constrained by the scope of the target variable in the dataset. The “risk level” classification provides a generalized assessment of maternal health risk but lacks clinical specificity regarding particular maternal health conditions. This limitation reduces the findings’ applicability in real-world clinical settings, where more nuanced and condition-specific risk factors are critical for decision-making. Additionally, the dataset lacks essential demographic and clinical details, such as distinctions between pregnancy stages, prenatal and postnatal complications, and other relevant health indicators. The inclusion of a broad age range from 10 to 70 years, extending beyond the typical reproductive age, introduces variability that may not accurately represent a maternal health population. This demographic inconsistency further limits the study’s relevance to clinical maternal health risk prediction.

Conclusions

In conclusion, this study serves as a demonstration of ML model applications to a maternal health dataset. The Decision Tree model proved the most effective, offering high accuracy and interpretability, while the KNN model also showed strong performance as a simpler alternative. While the SVM model has potential for non-linear data, it requires careful optimization for better performance. Logistic regression serves as a valuable baseline model, although it lacks capacity for handling complex data patterns. Future work could explore ensemble methods, such as Random Forest or boosting, or incorporate more advanced feature engineering techniques to further enhance the performance and generalizability of these models in maternal health risk prediction. However, we emphasize that the findings of this study are exploratory and not intended for clinical application without validation from subject-matter experts and comprehensive datasets specifically tailored to maternal health.

Acknowledgments

None.

Footnote

Peer Review File: Avaliable at https://jmai.amegroups.com/article/view/10.21037/jmai-24-332/prf

Funding: None.

Conflicts of Interest: The author has completed the ICMJE uniform disclosure form (available at https://jmai.amegroups.com/article/view/10.21037/jmai-24-332/coif). The author has no conflicts of interest to declare.

Ethical Statement: The author is accountable for all aspects of the work in ensuring that questions related to the accuracy or integrity of any part of the work are appropriately investigated and resolved.

Open Access Statement: This is an Open Access article distributed in accordance with the Creative Commons Attribution-NonCommercial-NoDerivs 4.0 International License (CC BY-NC-ND 4.0), which permits the non-commercial replication and distribution of the article with the strict proviso that no changes or edits are made and the original work is properly cited (including links to both the formal publication through the relevant DOI and the license). See: https://creativecommons.org/licenses/by-nc-nd/4.0/.

References

1. Lubano K. Bridging the gap: Implementation science, knowledge translation, and innovation for advancing maternal health sustainable development goals. J Obstet Gynaecol East Cent Africa 2024. Available online: 10.59692/jogeca.v36i1.15810.59692/jogeca.v36i1.158
2. Ojong SA, Wamakima B, Moyer CA, et al. Maternal Health and Well-Being. In Oxford Research Encyclopedia of Global Public Health. Oxford, UK: Oxford University; 2023.
3. Lee H. Introducing Maternal Oral Health As Global Health And Public Health Agenda. Eur J Public Health 2022;32:ckac129.257.
4. Drife JO, Lewis G, Neilson JP, et al. International Maternal Health: Global Action. Why Mothers Died and How their Lives are Saved 2023:193-206.
5. Deardorff J, Tissue MM, Elliott P, et al. The Critical Value of Maternal and Child Health (MCH) to Graduate Training in Public Health: A Framework to Guide Education, Research and Practice. Matern Child Health J 2022;26:121-8. [Crossref] [PubMed]
6. Sajedinejad S, Majdzadeh R, Vedadhir A, et al. Maternal mortality: a cross-sectional study in global health. Global Health 2015;11:4. [Crossref] [PubMed]
7. Mu C, Yan Z, Zhu Y. Prediction of Maternal Health Risk based on Physiological Indicators. ACM International Conference Proceeding Series 2023:578-84.
8. Raza A, Siddiqui HUR, Munir K, et al. Ensemble learning-based feature engineering to analyze maternal health during pregnancy and health risk prediction. PLoS One 2022;17:e0276525. [Crossref] [PubMed]
9. Yang Y, Wu N. Gestational Diabetes Mellitus and Preeclampsia: Correlation and Influencing Factors. Front Cardiovasc Med 2022;9:831297. [Crossref] [PubMed]
10. Sandsæter HL, Horn J, Rich-Edwards JW, et al. Preeclampsia, gestational diabetes and later risk of cardiovascular disease: Women's experiences and motivation for lifestyle changes explored in focus group interviews. BMC Pregnancy Childbirth 2019;19:448. [Crossref] [PubMed]
11. Khanlou N, Haque N, Skinner A, et al. Scoping Review on Maternal Health among Immigrant and Refugee Women in Canada: Prenatal, Intrapartum, and Postnatal Care. J Pregnancy 2017;2017:8783294. [Crossref] [PubMed]
12. Phillimore J. Delivering maternity services in an era of superdiversity: the challenges of novelty and newness. Ethnic and Racial Studies 2014;38:568-82. [Crossref]
13. Arshad H, Tayyab M, Bilal M, et al. Trends and challenges in harnessing big data intelligence for health care transformation. Artificial Intelligence for Intelligent Systems: Fundamentals, Challenges, and Applications 2024:220-40.
14. Basile LJ, Carbonara N, Pellegrino R, et al. Business intelligence in the healthcare industry: The utilization of a data-driven approach to support clinical decision making. Technovation 2023;120:102482. [Crossref]
15. Zitnik M, Nguyen F, Wang B, et al. Machine Learning for Integrating Data in Biology and Medicine: Principles, Practice, and Opportunities. Inf Fusion 2019;50:71-91. [Crossref] [PubMed]
16. Betts KS, Kisely S, Alati R. Predicting common maternal postpartum complications: leveraging health administrative data and machine learning. BJOG 2019;126:702-9. [Crossref] [PubMed]
17. Alanazi A. Using machine learning for healthcare challenges and opportunities. Inform Med Unlocked 2022;30:100924. [Crossref]
18. Singh P, Singh SP, Singh DS. An introduction and review on machine learning applications in medicine and healthcare. 2019 IEEE Conference on Information and Communication Technology. doi: 10.1109/CICT48419.2019.9066250.10.1109/CICT48419.2019.9066250
19. Ferdous M, Debnath J, Chakraborty NR. Machine Learning Algorithms in Healthcare: A Literature Survey. 2020 11th International Conference on Computing, Communication and Networking Technologies, ICCCNT 2020.
20. Olusanya MO, Ogunsakin RE, Ghai M, et al. Accuracy of Machine Learning Classification Models for the Prediction of Type 2 Diabetes Mellitus: A Systematic Survey and Meta-Analysis Approach. Int J Environ Res Public Health 2022;19:14280. [Crossref] [PubMed]
21. Signorini MG, Pini N, Malovini A, et al. Integrating machine learning techniques and physiology based heart rate features for antepartum fetal monitoring. Comput Methods Programs Biomed 2020;185:105015. [Crossref] [PubMed]
22. Garcia-Canadilla P, Sanchez-Martinez S, Crispi F, et al. Machine Learning in Fetal Cardiology: What to Expect. Fetal Diagn Ther 2020;47:363-72. [Crossref] [PubMed]
23. Office of the Surgeon General (OSG). The Surgeon General’s Call to Action to Improve Maternal Health [Internet]. Washington (DC): US Department of Health and Human Services 2020. Strategies and actions: improving maternal health and reducing maternal mortality and morbidity. Available online: https://www.ncbi.nlm.nih.gov/books/NBK568218/
24. Dasari P, Garg S, Keerthana K. Maternal Health Indicators. Current Practice in Medical Science 2022;1:1-17.
25. Bauserman M, Thorsten VR, Nolen TL, et al. Maternal mortality in six low and lower-middle income countries from 2010 to 2018: risk factors and trends. Reprod Health 2020;17:173. [Crossref] [PubMed]
26. Jadhav S, Tirankar V, Jadhav C, et al. Maternal near miss - indicator of maternal health. Int J Sci Res 2023;1-3. [Crossref]
27. Gazeley U, Polizzi A, Romero-Prieto JE, et al. Lifetime risk of maternal near miss morbidity: a novel indicator of maternal health. Int J Epidemiol 2024;53:dyad169. [Crossref] [PubMed]
28. Akor L, Dajo U, Anthony-Ekpe H. The Impact of Socio-demographic Variables on Maternal Healthcare Accessibility among Women of Reproductive Age in Zamfara State, Nigeria. Asian Journal of Agricultural Extension Economics and Sociology 2024;42:404-16.
29. Minopoli M, Noël L, Dagge A, et al. Maternal ethnicity and socioeconomic deprivation: influence on adverse pregnancy outcomes. Ultrasound Obstet Gynecol 2024;64:187-92. [Crossref] [PubMed]
30. Alfred MC, Wilson D. Adverse Events in Maternal Care: Investigating Racial/Ethnic Disparities at the System Level. Proceedings of the International Symposium of Human Factors and Ergonomics in Healthcare 2022;11:70-5.
31. Shen JJ, Mojtahedi Z, Vanderlaan J, et al. Disparities in Adverse Maternal Outcomes Among Five Race and Ethnicity Groups. J Womens Health (Larchmt) 2022;31:1432-9. [Crossref] [PubMed]
32. Ajoseh SM, Raji SO. Socio-Cultural Correlates of Maternal Morbidity in Lagos State, Nigeria. Nigerian Journal of Sociology and Anthropology 2022;20:119-43. [Crossref]
33. Mutlu HB, Durmaz F, Yücel N, et al. Prediction of Maternal Health Risk with Traditional Machine Learning Methods. NATURENGS MTU Journal of Engineering and Natural Sciences Malatya Turgut Ozal University 2023;4:16-23. [Crossref]
34. Raihen MN, Akter S. Comparative Assessment of Several Effective Machine Learning Classification Methods for Maternal Health Risk. Computational Journal of Mathematical and Statistical Sciences 2024;3:161-76. [Crossref]
35. Mustamin NF, Aziz F, Firmansyah F, et al. Classification of Maternal Health Risk Using Three Models Naive Bayes Method. IJCCS 2023;17:395. [Crossref]
36. Thakkar D, Gandhi VC, Trivedi D. Forecasting Maternal Women's Health Risks using Random Forest Classifier. 7th International Conference on Inventive Computation Technologies, ICICT 2024:961-5.
37. Ukrit MF, Jeyavathana RB, Rani AL, et al. Maternal Health Risk Prediction with Machine Learning Methods. 2nd International Conference on Emerging Trends in Information Technology and Engineering, IC-ETITE 2024:1-9.
38. Sinambela DP, Rahmatullah B, Lah NHC, et al. Machine learning approaches for predicting postpartum hemorrhage: a comprehensive systematic literature review. Indonesian Journal of Electrical Engineering and Computer Science 2024;34:2087-95. [Crossref]
39. Yang X. Prediction of maternal health risk levels based on multiple machine learning algorithms. Applied and Computational Engineering 2024;74:142-7. [Crossref]
40. Goyal P, Suyal M. A Review on Analysis of K-Nearest Neighbor Classification Machine Learning Algorithms based on Supervised Learning. Article in International Journal of Engineering Trends and Technology 2022;70:43-8. [Crossref]
41. Pisner DA, Schnyer DM. Support vector machine. Machine Learning: Methods and Applications to Brain Disorders 2020:101-21.
42. Jamel L, Umer M, Saidani O, et al. Improving prediction of maternal health risks using PCA features and TreeNet model. PeerJ Comput Sci 2024;10:e1982. [Crossref] [PubMed]
43. Pawar L, Malhotra J, Sharma A, et al. A Robust Machine Learning Predictive Model for Maternal Health Risk. 3rd International Conference on Electronics and Sustainable Communication Systems, ICESC 2022 - Proceedings 2022:882-8.
44. Kumar N, Hoque MA, Shahjaman M, et al. Metabolomic Biomarker Identification in Presence of Outliers and Missing Values. Biomed Res Int 2017;2017:2437608. [Crossref] [PubMed]
45. Hancock JT, Khoshgoftaar TM. Survey on categorical data for neural networks. J Big Data 2020;7:1-41. [Crossref]
46. Cho K, Yoon T, Park S, et al. Using standardization for fair data evaluation. International Conference on Advanced Communication Technology, ICACT 2018;2018-February:585-9.
47. Cheng Q, Varshney PK, Arora MK. Logistic regression for feature selection and soft classification of remote sensing data. IEEE Geoscience and Remote Sensing Letters 2006;3:491-4. [Crossref]
48. Nguyen QH, Ly HB, Ho LS, et al. Influence of Data Splitting on Performance of Machine Learning Models in Prediction of Shear Strength of Soil. Math Probl Eng 2021;2021:4832864. [Crossref]
49. Schober P, Vetter TR. Logistic Regression in Medical Research. Anesth Analg 2021;132:365-6. [Crossref] [PubMed]
50. Zhang S. Challenges in KNN Classification. IEEE Trans Knowl Data Eng 2022;34:4663-75. [Crossref]
51. Suthaharan S. Support Vector Machine. In: Machine Learning Models and Algorithms for Big Data Classification. Integrated Series in Information Systems 2016:207-35.
52. Heydarian M, Doyle TE, Samavi R. MLCM: Multi-Label Confusion Matrix. IEEE Access 2022;10:19083-95.
53. Adnan M, Alarood AAS, Uddin MI, et al. Utilizing grid search cross-validation with adaptive boosting for augmenting performance of machine learning models. PeerJ Comput Sci 2022;8:e803. [Crossref] [PubMed]
54. Boateng EY, Otoo J, Abaye DA, et al. Basic Tenets of Classification Algorithms K-Nearest-Neighbor, Support Vector Machine, Random Forest and Neural Network: A Review. Journal of Data Analysis and Information Processing 2020;8:341-57. [Crossref]
55. Zhou ZH. Support Vector Machine. Mach Learn 2021;129-53. [PubMed]
56. Nattino G, Pennell ML, Lemeshow S. Assessing the goodness of fit of logistic regression models in large samples: A modification of the Hosmer-Lemeshow test. Biometrics 2020;76:549-60. [Crossref] [PubMed]