Scholarlink

<p>Maternal morbidity and mortality remain major public health challenges, particularly in settings where antenatal triage is constrained by high demand and limited specialist availability. This study evaluates supervised machine learning for three-level gestational risk stratification (low, mid, and high risk) using routinely collected physiological measurements, focusing on class-specific errors, stability, and interpretability. A unified dataset was constructed by integrating two public maternal risk datasets from Bangladesh (N = 1,014) and Tanzania (N = 559 after preprocessing), including variable alignment, unit normalization, and transformation of longitudinal records into cross-sectional representations. Six classifiers (Multilayer Perceptron, Logistic Regression, Decision Tree, Random Forest, XGBoost, and Support Vector Machine) were evaluated using stratified 5-fold cross-validation. Without SMOTE, performance was asymmetric: low- and high-risk classes were well discriminated (F1 = 0.79-0.86 and 0.73-0.84), while the intermediate class showed higher uncertainty (F1 = 0.38-0.50), with macro-F1 ranging from 0.6684 to 0.7125. With SMOTE applied to training folds, performance improved, particularly for the intermediate class (F1 = 0.66-0.70), increasing macro-F1 to approximately 0.82-0.85. SHAP analysis identified blood glucose and systolic blood pressure as the most influential predictors. To demonstrate pipeline integration, a platform-independent prototype decision-support interface was developed, enabling structured data input, model inference, and visualization of predictions and feature attributions. The tool provides risk classification with explanatory outputs but was not implemented or evaluated in a real healthcare setting. Thus, it should be interpreted as a methodological proof of concept rather than a clinically validated system. Further work is required, including external validation, prospective evaluation, and integration into healthcare workflows. Overall, the results show that machine learning models effectively discriminate well-defined risk groups, while uncertainty persists in intermediate-risk cases, supporting future development of decision-support tools in maternal health.</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

Implementation-oriented machine learning for three-level maternal risk triage in primary antenatal care: stability, error profiles, and SHAP-based explainability

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年7月1日

查看全文
  • Fernando Rodrigues Trindade Ferreira,
  • Loena Marins do Couto,
  • Leticia Bussinger das Neves,
  • Guilherme de Melo Baptista Domingues

摘要

Maternal morbidity and mortality remain major public health challenges, particularly in settings where antenatal triage is constrained by high demand and limited specialist availability. This study evaluates supervised machine learning for three-level gestational risk stratification (low, mid, and high risk) using routinely collected physiological measurements, focusing on class-specific errors, stability, and interpretability. A unified dataset was constructed by integrating two public maternal risk datasets from Bangladesh (N = 1,014) and Tanzania (N = 559 after preprocessing), including variable alignment, unit normalization, and transformation of longitudinal records into cross-sectional representations. Six classifiers (Multilayer Perceptron, Logistic Regression, Decision Tree, Random Forest, XGBoost, and Support Vector Machine) were evaluated using stratified 5-fold cross-validation. Without SMOTE, performance was asymmetric: low- and high-risk classes were well discriminated (F1 = 0.79-0.86 and 0.73-0.84), while the intermediate class showed higher uncertainty (F1 = 0.38-0.50), with macro-F1 ranging from 0.6684 to 0.7125. With SMOTE applied to training folds, performance improved, particularly for the intermediate class (F1 = 0.66-0.70), increasing macro-F1 to approximately 0.82-0.85. SHAP analysis identified blood glucose and systolic blood pressure as the most influential predictors. To demonstrate pipeline integration, a platform-independent prototype decision-support interface was developed, enabling structured data input, model inference, and visualization of predictions and feature attributions. The tool provides risk classification with explanatory outputs but was not implemented or evaluated in a real healthcare setting. Thus, it should be interpreted as a methodological proof of concept rather than a clinically validated system. Further work is required, including external validation, prospective evaluation, and integration into healthcare workflows. Overall, the results show that machine learning models effectively discriminate well-defined risk groups, while uncertainty persists in intermediate-risk cases, supporting future development of decision-support tools in maternal health.