Scholarlink

Purpose <p>To develop a radiomics model utilizing contrast-enhanced computed tomography (CECT) for predicting the prognosis of post-acute pancreatitis diabetes mellitus (PPDM-A) and to explain the model’s internal predictive mechanisms using Shapley Additive exPlanations (SHAP).</p> Methods <p>226 PPDM-A patients were retrospectively recruited from three centers, with 107 in training, 46 in internal, and 73 in external cohorts. There were 34, 15 and 28 patients with complications in each cohort. The complications included microvascular complications, infection, diabetic ketosis and hypoglycemia. In PPDM-A patients’ first pancreatitis episode, 2398 radiomics features were extracted from CECT images (arterial and venous phases). FeAture Explorer generated the machine learning pipeline and selected important radiomics features. Gaussian processes classifiers built radiomics and clinical-radiomics models, while Naive Bayes classifiers built the clinical model. The SHAP method was applied to provide insights into the model’s predictive process.</p> Results <p>The radiomics model predicted PPDM-A complications with the area under the receiver operating characteristic curve (AUC) of 0.95, 0.888, and 0.948 in training, internal, and external cohorts. In external cohort, the radiomics model significantly outperformed the clinical model (AUC 0.948 vs. 0.713, <i>p</i> = 0.002), while the combined model showed no significant difference from the radiomics model (AUC 0.933 vs. 0.948, <i>p</i> = 0.638). The SHAP technology provided physicians with insights into the global and individual impacts of radiomic features on model predictions.</p> Conclusion <p>The CECT‑based radiomics model showed favorable prognostic performance for the prognosis of PPDM‑A. SHAP analysis interpreted the model’s mechanism, enhancing its clinical reliability and transparency.</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

An interpretable radiomics model based on contrast‑enhanced pancreatic computed tomography for predicting the prognosis of post-acute pancreatitis diabetes mellitus

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年3月13日

查看全文
  • Ran Hu,
  • Yan-Li Chen,
  • Gang-Jing Li,
  • Yin-Deng Luo,
  • Di Zhou,
  • Zhi-Gang Wang,
  • Xiao-Di Zhang,
  • Zhixuan Song,
  • Wei Chen,
  • Hua Yang

摘要

Purpose

To develop a radiomics model utilizing contrast-enhanced computed tomography (CECT) for predicting the prognosis of post-acute pancreatitis diabetes mellitus (PPDM-A) and to explain the model’s internal predictive mechanisms using Shapley Additive exPlanations (SHAP).

Methods

226 PPDM-A patients were retrospectively recruited from three centers, with 107 in training, 46 in internal, and 73 in external cohorts. There were 34, 15 and 28 patients with complications in each cohort. The complications included microvascular complications, infection, diabetic ketosis and hypoglycemia. In PPDM-A patients’ first pancreatitis episode, 2398 radiomics features were extracted from CECT images (arterial and venous phases). FeAture Explorer generated the machine learning pipeline and selected important radiomics features. Gaussian processes classifiers built radiomics and clinical-radiomics models, while Naive Bayes classifiers built the clinical model. The SHAP method was applied to provide insights into the model’s predictive process.

Results

The radiomics model predicted PPDM-A complications with the area under the receiver operating characteristic curve (AUC) of 0.95, 0.888, and 0.948 in training, internal, and external cohorts. In external cohort, the radiomics model significantly outperformed the clinical model (AUC 0.948 vs. 0.713, p = 0.002), while the combined model showed no significant difference from the radiomics model (AUC 0.933 vs. 0.948, p = 0.638). The SHAP technology provided physicians with insights into the global and individual impacts of radiomic features on model predictions.

Conclusion

The CECT‑based radiomics model showed favorable prognostic performance for the prognosis of PPDM‑A. SHAP analysis interpreted the model’s mechanism, enhancing its clinical reliability and transparency.