Scholarlink

<p>Accurate differentiation between benign and malignant thyroid nodules remains a persistent challenge in endocrine oncology due to overlapping clinical and sonographic characteristics. Traditional diagnostic tools such as Fine-Needle Aspiration Cytology (FNAC) frequently generate indeterminate outcomes, contributing to diagnostic ambiguity and resulting in a significant number of avoidable surgical procedures. To address these limitations, this study presents an interpretable and performance-enhanced ensemble learning approach named the Moderate Gradient Boost Machine (MGBM). The proposed framework integrates three key components: an evolutionary Remora Optimization Algorithm (ROA) for selecting the most informative features, the Synthetic Minority Over-sampling Technique (SMOTE) for mitigating class imbalance, and SHapley Additive exPlanations (SHAP) to ensure model transparency and clinical interpretability. Using a curated dataset of 1,232 thyroid cases, the MGBM-ROA-SHAP model demonstrated superior predictive capability, achieving 94.81% accuracy, 97.17% precision, 92.31% recall, 94.68% F1-score, and an exceptional ROC-AUC of 99.36%, outperforming established ensemble baselines. SHAP-based interpretability analysis identified TGAb levels, calcification, nodule size, and multifocality as the most influential malignancy indicators, offering clinically aligned and pathophysiologically meaningful explanations. These findings confirm that fusing evolutionary feature optimization with interpretable ensemble learning yields a powerful, transparent, and reliable diagnostic framework for early and accurate thyroid cancer prediction.</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

Interpretable MGBM framework with evolutionary feature optimization for robust thyroid malignancy prediction

  • 期刊论文

  • 发表日期：

    2026年6月18日

查看全文
  • Mohammed Yacoob B A,
  • Jayashree J

摘要

Accurate differentiation between benign and malignant thyroid nodules remains a persistent challenge in endocrine oncology due to overlapping clinical and sonographic characteristics. Traditional diagnostic tools such as Fine-Needle Aspiration Cytology (FNAC) frequently generate indeterminate outcomes, contributing to diagnostic ambiguity and resulting in a significant number of avoidable surgical procedures. To address these limitations, this study presents an interpretable and performance-enhanced ensemble learning approach named the Moderate Gradient Boost Machine (MGBM). The proposed framework integrates three key components: an evolutionary Remora Optimization Algorithm (ROA) for selecting the most informative features, the Synthetic Minority Over-sampling Technique (SMOTE) for mitigating class imbalance, and SHapley Additive exPlanations (SHAP) to ensure model transparency and clinical interpretability. Using a curated dataset of 1,232 thyroid cases, the MGBM-ROA-SHAP model demonstrated superior predictive capability, achieving 94.81% accuracy, 97.17% precision, 92.31% recall, 94.68% F1-score, and an exceptional ROC-AUC of 99.36%, outperforming established ensemble baselines. SHAP-based interpretability analysis identified TGAb levels, calcification, nodule size, and multifocality as the most influential malignancy indicators, offering clinically aligned and pathophysiologically meaningful explanations. These findings confirm that fusing evolutionary feature optimization with interpretable ensemble learning yields a powerful, transparent, and reliable diagnostic framework for early and accurate thyroid cancer prediction.