Introduction

Maternal health is crucial for ensuring healthy pregnancies and live births. It remains a major health priority, especially in developing countries, where 95% of preventable maternal deaths occur [1]. According to the World Health Organization, an estimated 86% (254,000) of global maternal deaths in 2017 occurred in Sub-Saharan Africa and Southern Asia. Specifically, two-thirds (196,000) of these deaths were in Sub-Saharan Africa, while one-fifth (58,000) occurred in Southern Asia [2].

Every pregnant woman hopes for a safe delivery, avoiding the loss of her life or her baby’s life. However, pregnancy outcomes such as live birth, spontaneous abortion, preterm delivery, or stillbirth are not always predictable by expectant mothers [3]. Approximately 800 women die each day due to childbirth-related complications. Women in developing countries are 36 times more likely to experience pregnancy-related complications compared to those in developed countries [4].

Pregnancies can be terminated safely using WHO-recommended methods, such as medical abortion with medications like mifepristone and misoprostol, or through manual vacuum aspiration (MVA) or dilation and curettage (D&C) performed by trained healthcare providers in appropriate medical settings [5]. Alternatively, pregnancies can be terminated unsafely through self-induced abortions, procedures in unhygienic conditions, or by non-medical practitioners. Globally, Six out of ten unintended pregnancies result in induced abortions, with 97% of unsafe abortions representing about 45% of all abortions occurring in developing nations. Unsafe abortions account for 4.7% to 13.2% of maternal deaths annually. It is estimated that 30 women in developed countries undergo unsafe abortions per 100,000 live births [6].

Preterm birth, a common complication, is the leading cause of death in children under five. Each year, an estimated 15 million babies are born preterm (before 37 weeks of gestation), with preterm birth rates ranging from 5% to 18% of all births, depending on the country. In low-income countries, more than 90% of extremely preterm newborns (born before 28 weeks) die within the first few days of life [7].

Stillbirth occurs when a baby dies after 28 weeks of pregnancy but before or during birth. Nearly 2 million stillbirths occur each year, equating to one every 16 seconds. According to the report "A Neglected Tragedy: The Global Burden of Stillbirths," 84% of these cases occur in low- and lower-middle-income countries. In 2019, three out of four stillbirths occurred in Sub-Saharan Africa or Southern Asia [8].

According to Rwanda’s recent demographic and health survey [9], the perinatal mortality rate was 28 deaths per 1,000 pregnancies. This rate includes early neonatal deaths (live births that do not survive past the first 7 days after birth) and stillbirths (pregnancy losses occurring after 7 months of gestation).

Machine learning (ML) has been extensively used to identify potential maternal risks during pregnancy and to forecast childbirth outcomes [10]. In this study, ML algorithms were employed to predict pregnancy outcomes in Rwanda using data from the recent Rwandan demographic and health survey [9]. ML has diverse applications across various fields, including classification and prediction (e.g., pandemic forecasting and medical diagnosis), clustering analysis (e.g., web security through unusual traffic detection, cancer cell identification, and customer segmentation), and natural language processing (e.g., speech recognition, language translation, and sentiment analysis).

Unlike traditional methods that often rely on linear assumptions and limited variable interactions, ML algorithms can handle complex, non-linear relationships and large datasets with numerous predictors, this allows for more accurate and nuanced predictions of pregnancy outcomes [11]. While some researchers have explored forecasting potential pregnancy complications and identifying the optimal delivery method based on maternal features [12], ML has yet to be applied to predict pregnancy outcomes among reproductive women in Rwanda.

Materials and methods

Results

Social demographic characteristics of the respondents

The sociodemographic and maternal characteristics of the respondents were analyzed using descriptive statistics, frequency tables and percentage distribution for each variable are presented in (Table 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 1. Demographic and maternal characteristics of reproductive women in Rwanda.

https://doi.org/10.1371/journal.pone.0312447.t001

The results from the 2019–2020 Demographic and Health Survey revealed that among the 14,634 reproductive women surveyed, approximately 22.6% were aged between 15 and 19 years, while only 8.4% were aged 45 to 49 years. Regarding education, 9.2% of respondents had never attended school, more than half (58.1%) had completed primary education, and only 4.6% had received tertiary education. Most women had never been in union (41.4%), 32.2% were married and 17.7% were in informal unions. The participants belonged to various religious groups, with approximately 46.2% identifying as Protestants, 37.6% as Catholics, and only a small number practicing traditional beliefs (1).

In terms of residential areas, 75.7% of women lived in rural regions. Employment data showed that more than half (55.9%) were engaged in the services sector, 27% were unemployed at the time of the survey, and 10.1% worked in unskilled manual labor. Wealth status varied, with 23.8% from the wealthiest families, 19.7% from wealthier families, and 18.5% from poorer families. Regarding access to healthcare, 79% of women reported that reaching a health facility was not a significant problem. In terms of pregnancy planning, 96% of women desired to have children later, while 3.3% wanted to become pregnant immediately.

On average, women had their first sexual intercourse at age 14. Healthcare provider visits were reported by 61.1% of women before the survey period. During their pregnancies, 45.4% attended all four recommended antenatal care (ANC) visits, 2.3% attended more than four visits, and 2.1% did not attend any ANC visits. Smoking was reported by only 0.6% of respondents. Women had been in an average of 1.3 unions, and 15.1% of participants had experienced sexual violence

The pregnancy outcomes among reproductive women in Rwanda

The results indicated that most women who participated in the recent demographic and health surveys delivered live births (93.4%). However, 4.5% of pregnancies ended in miscarriage, and 2.1% resulted in stillbirths (Fig 1).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 1. Pregnancy outcomes among reproductive women in Rwanda.

https://doi.org/10.1371/journal.pone.0312447.g001

Risk factors associated with adverse pregnancy outcome among reproductive women in Rwanda

Bivariate analysis.

The association between independent variables and diverse pregnancy outcomes (the dependent variable) was evaluated. Both chi-square statistics and p-values were computed, with a p-value less than 0.05 indicating a statistically significant association (Table 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 2. Bivariate analysis: Factors associated with adverse pregnancy outcome among reproductive women in Rwanda.

https://doi.org/10.1371/journal.pone.0312447.t002

The findings revealed that most women, regardless of age, delivered live births. However, a higher proportion of adverse pregnancy outcomes was observed among women aged 35 years and older. Specifically, 20.4% of women aged 35–39, 25.3% of those aged 40–44, and 27.2% of women aged 45–49 experienced adverse pregnancy outcomes. Age was significantly associated with adverse pregnancy outcomes (p < 0.001).

Educational level was not significantly associated with adverse pregnancy outcomes (p = 0.106), although a higher proportion of adverse outcomes among women with no formal education (18.8%). Marital status was significantly linked to adverse pregnancy outcomes (p < 0.001), with married women experiencing more adverse outcomes (22.8%). Religion was not associated with adverse pregnancy outcomes (p = 0.600), with a higher proportion observed among women with no religious affiliation (18.9%), followed by those in Protestant denominations (12.9%).

Residential location was not associated with adverse pregnancy outcomes (p = 0.894), with 12.4% of women in rural areas experiencing adverse outcomes compared to 11.1% in urban areas. The gender of the household head also not correlated with adverse pregnancy outcomes (p = 0.331), with male-headed households showing a higher rate of adverse outcomes (13.5%). The occupation of respondents was not associated with adverse pregnancy outcomes (p = 0.062), with a higher proportion observed among women working in the services sector (16.7%) compared to those in skilled manual jobs (3.8%).

Distance to health facilities was not linked to adverse pregnancy outcomes (p = 0.205), despite a higher proportion of adverse pregnancies among women who reported difficulty accessing health facilities (13.5%). Pregnancy desire was significantly associated with adverse outcomes (p < 0.001), with more adverse pregnancies observed among women who desired pregnancy (21.7%). Age at first sexual intercourse was significantly associated with adverse pregnancy outcomes (p < 0.001), with women experiencing adverse outcomes starting their sexual activity later (average age 20.5 years) compared to those who started earlier (average age 13.6 years).

Being visited by a health care provider was significantly associated with adverse pregnancy outcomes (p < 0.001). ANC utilization was also significant (p = 0.002), with a higher proportion of adverse outcomes observed among women who attended more than four ANC visits (18.1%). Finally, the number of unions a woman had been in was significantly associated with adverse pregnancy outcomes (p < 0.001).

Multivariable analysis of factors associated with adverse pregnancy outcome among reproductive women in Rwanda

To explore risk factors while adjusting for confounders and effect modification, binary logistic regression was performed. The analysis calculated the adjusted odds ratios (AOR), 95% confidence intervals, and p-values for variables that demonstrated significant relationships in the bivariate analysis. This approach allowed for a more nuanced understanding of the associations between predictors and adverse pregnancy outcomes, accounting for potential confounding variables and interactions (Table 3).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 3. Multivariate analysis of risk factors associated with adverse pregnancy outcome among reproductive women in Rwanda.

https://doi.org/10.1371/journal.pone.0312447.t003

The findings revealed that the respondents’ age was significantly associated with adverse pregnancy outcomes. As age increased, so did the risk of adverse pregnancy outcomes. For women aged 30–34 years, the risk was more than 5 times higher (AOR = 5.755, 95% CI = 3.085–10.074, p < 0.001), and for those aged 35–39 years, it was more than 8 times higher (AOR = 8.458, 95% CI = 4.507–10.571, p < 0.001). The risk was around 12 times higher for women aged 40–44 years (AOR = 11.86, 95% CI = 6.250–21.842, p < 0.001), and more than 14 times higher for those aged 45–49 years (AOR = 14.233, 95% CI = 7.359–25.922, p < 0.001), compared to women aged 15–19 years.

Marital status also significantly influenced adverse pregnancy outcomes. Married women were 0.9 times less likely to experience adverse outcomes compared to their unmarried counterparts (AOR = 0.894, 95% CI = 0.787–0.966, p = 0.011). The women who were not visited by a healthcare provider during pregnancy were more likely to experience advserse pregnancy outcome (AOR = 1.421, 95% CI = 1.300–1.611).

Antenatal care (ANC) utilization showed a protective effect. Women who attended at least two ANC visits were 0.8 times less likely to experience adverse pregnancy outcomes compared to those who did not attend any ANC visits (AOR = 0.801, 95% CI = 0.664–0.9615, p = 0.019). Finally, the number of unions a woman had experienced was positively associated with adverse pregnancy outcomes. Each additional union increased the likelihood of adverse outcomes by 1.32 times (AOR = 1.320; 95% CI = 1.104–1.573; p = 0.002).

Model performance evaluation

The K-Nearest Neighbors (KNN) model demonstrated a strong performance with a recall score of 0.97, indicating its high ability to correctly identify most instances of adverse pregnancy outcomes. Its precision score of 0.89 also reflects a good capability in correctly identifying positive cases. The AUC score of 0.842 for KNN suggests that it is highly capable of distinguishing between different outcome categories, making it one of the top-performing models.

The Decision Tree model exhibited a recall score of 0.95 and a precision score of 0.88. This shows that it is effective at identifying adverse pregnancy outcomes and accurately classifying positive instances. Despite these strong scores, the Decision Tree’s F1 and AUC scores were slightly lower than those of the KNN model and the Random Forest, indicating that its overall performance might be less robust. The Random Forest model achieved an AUC score of 0.81, which, while higher than that of the Naïve Bayes model, was lower than the AUC scores for Logistic Regression, Support Vector Machine (SVM), Neural Network, and KNN models. Despite this, the Random Forest model’s accuracy, precision, recall, and F1 scores were commendable, underscoring its effectiveness in predicting adverse pregnancy outcomes, though it did not reach the highest performance levels of the top models.

The KNN model highlighted that woman with advanced maternal age (>30 years), those in multiple unions, and those who were not visited by a health care provider were more prone to adverse pregnancy outcomes. Additionally, women who lacked adequate antenatal care (ANC) visits and unmarried women were identified as being at higher risk. These findings are consistent with the results from the logistic regression model, further validating the predictive power of machine learning in identifying high-risk categories for adverse pregnancy outcomes.

Overall, the KNN, Decision Tree, and Random Forest models demonstrated high AUC scores relative to other models, marking them as effective tools for predicting adverse pregnancy outcomes (Table 4 and Fig 2).

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Fig 2. Overall model performance.

https://doi.org/10.1371/journal.pone.0312447.g002

Download:

• PPT
  PowerPoint slide
• PNG
  larger image
• TIFF
  original image

Table 4. Overall model performance.

https://doi.org/10.1371/journal.pone.0312447.t004

Discussion

The results from our study revealed that, during the five years preceding the 2019–2020 Rwanda Demographic and Health Survey, 93.4% of pregnancies resulted in live births, 4.5% ended in miscarriages, and 2.1% ended in stillbirths. These findings align with the recent Rwanda Demographic and Health Survey, which reported that 21 out of every 1,000 pregnancies end in stillbirths. Notably, the miscarriage rate in our study was 20% lower than the global estimate, indicating that Rwanda is making progress in reducing pregnancy loss rates through improvements in maternal health services [9].

Our analysis also highlighted that age was a significant factor associated with adverse pregnancy outcomes. As the age of respondents increased, so did the risk of experiencing adverse pregnancy outcomes. This finding is consistent with a study in China which has shown that pregnancy at the age of 35 years or older, known as advanced maternal age (>30 years), is a notable risk factor for negative pregnancy and childbirth outcomes compared to younger women [14].

Our study found that marital status significantly influenced adverse pregnancy outcomes. Compared to single women, married women were less likely to experience adverse pregnancy outcomes. This finding contrast with research conducted in the United States, which reported that unmarried mothers had a higher risk of infant mortality compared to married mothers, even when accounting for factors such as maternal age and education [15], the difference could be explained by differences in population demographics, healthcare access and cultural norms. Married women may benefit from increased social support and engagement in healthier behaviours, such as regular prenatal care, a nutritious diet, and exercise, which can mitigate the risk of adverse outcomes. In contrast, unmarried women may face additional stressors, including social stigma and lack of support, which can negatively impact pregnancy outcomes [16].

Our results also demonstrated that women who were visited by healthcare provider during pregnancy were less likely to encounter adverse pregnancy outcomes. Moreover, women who attended two or more antenatal care (ANC) visits were 0.8 times less likely to experience adverse pregnancy outcomes compared to those who did not attend any ANC visits. This finding aligns with a study conducted in Ghana, which indicated that regular attendance at ANC appointments and receiving appropriate prenatal care significantly improved pregnancy outcomes, including reduced risks of maternal and infant mortality, preterm birth, and low birth weight [17].

Additionally, our study revealed that the number of unions a woman has had was significantly associated with adverse pregnancy outcomes. Women with more unions were more likely to experience adverse pregnancy outcomes. This is supported by studies suggesting that women with multiple unions may face higher risks of preterm birth, low birth weight, and infant mortality compared to those with fewer unions [18]. Women with multiple unions might engage in behaviors that increase these risks, such as smoking, substance abuse, or risky sexual behavior, which can negatively affect both maternal and fetal health.

The overarching goal of this study was to predict adverse pregnancy outcomes using supervised machine learning algorithms. We employed and evaluated seven machine learning algorithms: decision trees, neural networks, random forests, K-nearest neighbors (KNN), naïve Bayes, logistic regression, and support vector machines (SVM). Among these, the KNN, decision tree, and random forest models demonstrated high AUC scores and were effective in predicting adverse pregnancy outcomes.

This study has significant implications for research, policy, and practice. In research, the findings demonstrate the potential of machine learning algorithms to predict adverse pregnancy outcomes, suggesting the need for future studies to validate these models using diverse datasets and real-time clinical data. In terms of policy, the study highlights the importance of strengthening maternal health programs, especially by expanding access to antenatal care (ANC) services and targeting high-risk groups such as older women and those with multiple unions. From a practice standpoint, healthcare providers can use machine learning tools to improve risk stratification and provide tailored care to pregnant women, ensuring early interventions and better maternal and neonatal outcomes.

Conclusion

The findings revealed that 4.5% of pregnancies ended in miscarriage and 2.1% in stillbirths. Key risk factors for adverse pregnancy outcomes included advanced maternal age (>30 years), being unmarried, no visit by a healthcare provider, inadequate antenatal care utilization, and having multiple unions. Among the machine learning models evaluated, the K-Nearest Neighbors (KNN) algorithm demonstrated superior performance, achieving the highest predictive accuracy for adverse pregnancy outcomes. While these results suggest that KNN is a promising tool for early risk identification, further research is needed to validate these findings across broader populations and settings before machine learning models can be integrated into clinical practice for maternal healthcare in Rwanda.

Limitation of the study

This study has some limitations. First, the cross-sectional design limits the ability to establish causality between identified risk factors and adverse pregnancy outcomes. Additionally, the use of self-reported data from the RDHS may lead to recall bias or inaccuracies. Our dependent variable, which categorized outcomes as either live birth or pregnancy loss, does not account for the complexity of whether the pregnancy was desired, pregnancy loss may not always be considered adverse in the case of an unwanted pregnancy. Furthermore, maternal mortality and live birth complications were not included in the analysis, which could provide a more comprehensive view of pregnancy outcomes. Lastly, while machine learning models demonstrated predictive capabilities, their performance could be improved with more detailed clinical data and external validation using independent datasets.

Study contribution

The study provides valuable insights into the use of machine learning techniques to predict adverse pregnancy outcomes. Offering an innovative approach compared to traditional statistical methods. By applying models likes K-nearest neighbors (KNN) and validating their performance, the study demonstrates how data-driven techniques can improve the accuracy of predictions related to maternal health, particularly in resource-limited settings. The study contributed to the understanding of the specific risk factors, such as advanced maternal age, marital status, and ANC attendance, in relation to adverse pregnancy outcomes in Rwanda. This context-specific analysis provides policy makers wand healthcare provides with actionable information that could help target interventions more effectively to reduce maternal and neonatal morbidity and mortality.

Acknowledgments

We would like to thank the African Center of Excellence in Data Science at the University of Rwanda for their enthusiasm and support. We are also deeply grateful to our families and friends for their encouragement throughout this work.