Scholarlink

Objective <p>To construct and validate an imaging-based predictive model grounded in multiscale progressive structural representation, and to systematically evaluate the incremental diagnostic value from whole-brain macroscopic gray matter volume to hippocampal subfield microstructural texture features in identifying mild cognitive impairment associated with type 2 diabetes mellitus (T2DM-MCI).</p> Methods <p>A total of 280 patients with T2DM who met the diagnostic criteria of the American Diabetes Association were retrospectively enrolled, including 82 patients with T2DM-MCI and 198 cognitively normal individuals. All participants underwent 3.0T structural MRI scanning. Whole-brain gray matter volume (GMV) features were extracted based on the AAL atlas. Bilateral 24 hippocampal subfield volumetric features were segmented using FreeSurfer, and 2,232 hippocampal subfield radiomic features were extracted using PyRadiomics. A nested cross-validation framework was constructed, and stepwise dimensionality reduction was performed using the Mann–Whitney U test, mRMR, and LASSO to develop the GMV model, hippocampal subfield volume model (Hip-Volume), and hippocampal subfield radiomics model (Hip-Radscore), respectively. A combined model integrating clinical variables was further established. Model performance was evaluated using AUC, sensitivity, specificity, PPV, and NPV, with comparisons conducted using the DeLong test. Clinical utility was assessed using NRI, IDI, Brier score, and decision curve analysis (DCA). Model explainability was achieved through SHAP analysis, and Spearman correlation was applied to examine the association between Hip-Radscore and MMSE.</p> Results <p>The whole-brain GMV model demonstrated limited discriminative performance (AUC = 0.63 ± 0.04). When the analysis scale was focused on hippocampal subfield volumes, model performance improved significantly (AUC = 0.71 ± 0.03, <i>P</i> &lt; 0.05). Under consistent anatomical regions, the hippocampal subfield radiomics model further increased to an AUC of 0.78 ± 0.03 (<i>P</i> = 0.004). The combined model integrating clinical variables and hippocampal subfield radiomic features achieved the best performance (AUC = 0.86 ± 0.02) and showed significantly superior reclassification ability compared with the “clinical + volume” model (NRI = 0.30, IDI = 0.08, both <i>P</i> &lt; 0.001). SHAP analysis identified Hip-Radscore as the most important predictor. Hip-Radscore was significantly negatively correlated with MMSE scores (ρ = −0.58, <i>P</i> &lt; 0.001).</p> Conclusion <p>Multiscale progressive structural representation significantly improves the identification performance of T2DM-MCI, and hippocampal subfield radiomic features demonstrate greater early sensitivity than traditional volumetric indicators.</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

Progressive modeling of multiscale hippocampal subfield structural representation improves the prediction performance of mild cognitive impairment in type 2 diabetes

  • 期刊论文

  • 发表日期：

    2026年4月1日

查看全文
  • Yu Shen,
  • Ran-chao Wang,
  • Jin-xin Wang,
  • Yang Li,
  • Yu-hao Xu,
  • Yue-feng Li

摘要

Objective

To construct and validate an imaging-based predictive model grounded in multiscale progressive structural representation, and to systematically evaluate the incremental diagnostic value from whole-brain macroscopic gray matter volume to hippocampal subfield microstructural texture features in identifying mild cognitive impairment associated with type 2 diabetes mellitus (T2DM-MCI).

Methods

A total of 280 patients with T2DM who met the diagnostic criteria of the American Diabetes Association were retrospectively enrolled, including 82 patients with T2DM-MCI and 198 cognitively normal individuals. All participants underwent 3.0T structural MRI scanning. Whole-brain gray matter volume (GMV) features were extracted based on the AAL atlas. Bilateral 24 hippocampal subfield volumetric features were segmented using FreeSurfer, and 2,232 hippocampal subfield radiomic features were extracted using PyRadiomics. A nested cross-validation framework was constructed, and stepwise dimensionality reduction was performed using the Mann–Whitney U test, mRMR, and LASSO to develop the GMV model, hippocampal subfield volume model (Hip-Volume), and hippocampal subfield radiomics model (Hip-Radscore), respectively. A combined model integrating clinical variables was further established. Model performance was evaluated using AUC, sensitivity, specificity, PPV, and NPV, with comparisons conducted using the DeLong test. Clinical utility was assessed using NRI, IDI, Brier score, and decision curve analysis (DCA). Model explainability was achieved through SHAP analysis, and Spearman correlation was applied to examine the association between Hip-Radscore and MMSE.

Results

The whole-brain GMV model demonstrated limited discriminative performance (AUC = 0.63 ± 0.04). When the analysis scale was focused on hippocampal subfield volumes, model performance improved significantly (AUC = 0.71 ± 0.03, P < 0.05). Under consistent anatomical regions, the hippocampal subfield radiomics model further increased to an AUC of 0.78 ± 0.03 (P = 0.004). The combined model integrating clinical variables and hippocampal subfield radiomic features achieved the best performance (AUC = 0.86 ± 0.02) and showed significantly superior reclassification ability compared with the “clinical + volume” model (NRI = 0.30, IDI = 0.08, both P < 0.001). SHAP analysis identified Hip-Radscore as the most important predictor. Hip-Radscore was significantly negatively correlated with MMSE scores (ρ = −0.58, P < 0.001).

Conclusion

Multiscale progressive structural representation significantly improves the identification performance of T2DM-MCI, and hippocampal subfield radiomic features demonstrate greater early sensitivity than traditional volumetric indicators.