Scholarlink

<p>Postpartum depression (PPD) is a major health concern impacting maternal and infant health, but it is underdiagnosed because of stigma, subjective screening instruments, and inadequacy of mental healthcare. This study proposes an interpretable machine learning (ML) framework for early PPD prediction, combining comprehensive feature engineering with Explainable AI (XAI) techniques to enhance both accuracy and clinical trustworthiness. A dataset comprising 410 instances and 47 clinical and demographic attributes was subjected to extensive preprocessing. Four distinct feature engineering strategies combining traditional statistical methods (rule-based association mining and voting ensemble selection) and XAI techniques (SHapley Additive exPlanations (SHAP)-based and Local Interpretable Model-agnostic Explanations (LIME)-based) were employed to systematically identify the most predictive features while enhancing interpretability and reducing dimensionality. A comparative evaluation of several ML algorithms and ensemble strategies was conducted across diverse feature configurations. An ablation study was conducted with model variations to evaluate the framework’s performance with and without its comprehensive feature engineering revealing significant improvement through SHAP-driven feature selection. A hyperparameter-tuned Random Forest model trained on 33 SHAP-selected features achieved state-of-the-art performance, with 94.1% accuracy. The model significantly outperformed other approaches, demonstrating value of XAI-driven feature reduction. Finally, the optimized model was integrated into a real-time web platform, enabling scalable, accessible PPD risk screening.</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

Transparent and tuned: XAI for postpartum depression screening with comprehensive feature engineering and real-time web-integration

  • 期刊论文

  • 发表日期：

    2026年4月2日

查看全文
  • G. M. Abdullah Al Kafi,
  • Sayma Alam Suha,
  • Golam Ali,
  • Md. Arafat Hossain,
  • Anirban Chanda Suhag

摘要

Postpartum depression (PPD) is a major health concern impacting maternal and infant health, but it is underdiagnosed because of stigma, subjective screening instruments, and inadequacy of mental healthcare. This study proposes an interpretable machine learning (ML) framework for early PPD prediction, combining comprehensive feature engineering with Explainable AI (XAI) techniques to enhance both accuracy and clinical trustworthiness. A dataset comprising 410 instances and 47 clinical and demographic attributes was subjected to extensive preprocessing. Four distinct feature engineering strategies combining traditional statistical methods (rule-based association mining and voting ensemble selection) and XAI techniques (SHapley Additive exPlanations (SHAP)-based and Local Interpretable Model-agnostic Explanations (LIME)-based) were employed to systematically identify the most predictive features while enhancing interpretability and reducing dimensionality. A comparative evaluation of several ML algorithms and ensemble strategies was conducted across diverse feature configurations. An ablation study was conducted with model variations to evaluate the framework’s performance with and without its comprehensive feature engineering revealing significant improvement through SHAP-driven feature selection. A hyperparameter-tuned Random Forest model trained on 33 SHAP-selected features achieved state-of-the-art performance, with 94.1% accuracy. The model significantly outperformed other approaches, demonstrating value of XAI-driven feature reduction. Finally, the optimized model was integrated into a real-time web platform, enabling scalable, accessible PPD risk screening.