Scholarlink

Background <p>Health facility delivery is a critical strategy for reducing maternal and neonatal mortalities. In Somalia, maternal mortality remains alarmingly high due to socioeconomic disparities, geographic barriers, and limited healthcare infrastructure. Machine learning (ML) offers a novel approach for predicting health facility deliveries and identifying key determinants, enabling targeted interventions to improve maternal health outcomes.</p> Methods <p>This study analyzed data from the 2020 Somalia Demographic and Health Survey (SDHS) involving 8,951 women aged 15–49&#xa0;years. Seven ML algorithms, Random Forest, XGBoost, Gradient Boosting, Logistic Regression, Support Vector Machine, Decision Tree, and K-Nearest Neighbors, were evaluated for their ability to predict health facility deliveries. Model performance was assessed using the accuracy, precision, recall, F1-score, and AUROC. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the relative importance of predictors, including wealth quintile, antenatal care (ANC) attendance, and residence type.</p> Results <p>The Random Forest model achieved the highest performance, with an accuracy of 82%, a recall of 84%, and an AUROC of 0.89. XGBoost and Gradient Boosting followed with accuracies of 80% and 77%, respectively, and AUROC values of 0.89 and 0.86. Logistic regression and support vector machines demonstrated moderate performance (accuracy: 72%, AUROC: 0.80–0.81). SHAP analysis identified the wealth quintile as the most influential predictor, with women in the highest quintile being six times more likely to deliver in health facilities than those in the lowest quintile (AOR: 6.01; 95% CI: 4.04–8.94). ANC attendance and residence type were also significant contributors, with women attending four or more ANC visits (AOR: 4.82; 95% CI: 3.75–6.20) and urban residents were more likely to deliver in health facilities.</p> Conclusion <p>Machine learning techniques, particularly Random Forest and SHAP analyses, offer robust tools for predicting health facility deliveries and identifying critical determinants. These findings underscore the potential of ML in designing targeted data-driven interventions to improve maternal health outcomes in Somalia.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Exploring the application of machine learning and SHAP explanations to predict health facility deliveries in Somalia

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2025年8月15日

查看全文
  • Jamilu Sani,
  • Salad Halane,
  • Mohamed Mustaf Ahmed,
  • Abdiwali Mohamed Ahmed,
  • Jamal Hassan Mohamoud

摘要

Background

Health facility delivery is a critical strategy for reducing maternal and neonatal mortalities. In Somalia, maternal mortality remains alarmingly high due to socioeconomic disparities, geographic barriers, and limited healthcare infrastructure. Machine learning (ML) offers a novel approach for predicting health facility deliveries and identifying key determinants, enabling targeted interventions to improve maternal health outcomes.

Methods

This study analyzed data from the 2020 Somalia Demographic and Health Survey (SDHS) involving 8,951 women aged 15–49 years. Seven ML algorithms, Random Forest, XGBoost, Gradient Boosting, Logistic Regression, Support Vector Machine, Decision Tree, and K-Nearest Neighbors, were evaluated for their ability to predict health facility deliveries. Model performance was assessed using the accuracy, precision, recall, F1-score, and AUROC. SHapley Additive exPlanations (SHAP) analysis was employed to interpret the relative importance of predictors, including wealth quintile, antenatal care (ANC) attendance, and residence type.

Results

The Random Forest model achieved the highest performance, with an accuracy of 82%, a recall of 84%, and an AUROC of 0.89. XGBoost and Gradient Boosting followed with accuracies of 80% and 77%, respectively, and AUROC values of 0.89 and 0.86. Logistic regression and support vector machines demonstrated moderate performance (accuracy: 72%, AUROC: 0.80–0.81). SHAP analysis identified the wealth quintile as the most influential predictor, with women in the highest quintile being six times more likely to deliver in health facilities than those in the lowest quintile (AOR: 6.01; 95% CI: 4.04–8.94). ANC attendance and residence type were also significant contributors, with women attending four or more ANC visits (AOR: 4.82; 95% CI: 3.75–6.20) and urban residents were more likely to deliver in health facilities.

Conclusion

Machine learning techniques, particularly Random Forest and SHAP analyses, offer robust tools for predicting health facility deliveries and identifying critical determinants. These findings underscore the potential of ML in designing targeted data-driven interventions to improve maternal health outcomes in Somalia.