Scholarlink

<p>As artificial intelligence (AI) becomes more common in important areas like healthcare, it is essential for models to be both accurate and easy for clinicians to understand. However, many powerful AI models work like ‘black boxes,’ making their decisions difficult to explain and limiting their use in clinical practice. This study investigates the integration of Explainable Artificial Intelligence (XAI) techniques into machine learning models for breast cancer risk prediction, aiming to balance the trade-off between model accuracy and interpretability to support safer, more transparent decision-making in clinical settings. A large-scale, real-world dataset from the Breast Cancer Surveillance Consortium (BCSC), comprising over 1.5 million records, was used. State-of-the-art machine learning algorithms, including Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and a model based on Dempster-Shafer Theory, were developed. To enhance explainability, SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), and Layer-wise Relevance Propagation (LRP) were applied. Model evaluation prioritized recall to minimize false negatives, a critical factor in cancer detection. The XGBoost model achieved the highest accuracy-recall trade-off but revealed some clinically questionable dependencies in feature importance. The SVM model, explained via LIME, aligned closely with known breast cancer risk factors, providing highly interpretable local explanations. The ANN model, though less performant, effectively leveraged LRP to highlight significant clinical predictors. The Dempster-Shafer-based classifier offered a strong compromise between interpretability and accuracy, delivering transparent decision rules alongside respectable performance metrics. Incorporating XAI methods into predictive models for breast cancer detection improves interpretability without substantial sacrifices in performance. The study underscores the importance of balancing accuracy with explainability in clinical AI applications and advocates for broader adoption of explainable models to foster transparency, trust, and ethical responsibility in healthcare AI systems.</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

Explainable AI for breast cancer risk prediction: evaluating the accuracy-explainability trade-off

  • 期刊论文

  • 发表日期：

    2025年6月28日

查看全文
  • Sona Barseghyan,
  • Armine Babajanyan,
  • Zhanna Balyan,
  • Anna Asatryan,
  • Sachin Kumar

摘要

As artificial intelligence (AI) becomes more common in important areas like healthcare, it is essential for models to be both accurate and easy for clinicians to understand. However, many powerful AI models work like ‘black boxes,’ making their decisions difficult to explain and limiting their use in clinical practice. This study investigates the integration of Explainable Artificial Intelligence (XAI) techniques into machine learning models for breast cancer risk prediction, aiming to balance the trade-off between model accuracy and interpretability to support safer, more transparent decision-making in clinical settings. A large-scale, real-world dataset from the Breast Cancer Surveillance Consortium (BCSC), comprising over 1.5 million records, was used. State-of-the-art machine learning algorithms, including Extreme Gradient Boosting (XGBoost), Support Vector Machines (SVM), Artificial Neural Networks (ANN), and a model based on Dempster-Shafer Theory, were developed. To enhance explainability, SHAP (Shapley Additive Explanations), LIME (Local Interpretable Model-Agnostic Explanations), and Layer-wise Relevance Propagation (LRP) were applied. Model evaluation prioritized recall to minimize false negatives, a critical factor in cancer detection. The XGBoost model achieved the highest accuracy-recall trade-off but revealed some clinically questionable dependencies in feature importance. The SVM model, explained via LIME, aligned closely with known breast cancer risk factors, providing highly interpretable local explanations. The ANN model, though less performant, effectively leveraged LRP to highlight significant clinical predictors. The Dempster-Shafer-based classifier offered a strong compromise between interpretability and accuracy, delivering transparent decision rules alongside respectable performance metrics. Incorporating XAI methods into predictive models for breast cancer detection improves interpretability without substantial sacrifices in performance. The study underscores the importance of balancing accuracy with explainability in clinical AI applications and advocates for broader adoption of explainable models to foster transparency, trust, and ethical responsibility in healthcare AI systems.