He Wenzhang, Cui Beibei, Chu Zhigang, Huang Xuan, Liu Jing, Li Xue, Wang Yinqiu, Pang Xueting, Lin Hui, Peng Liqing
Department of Radiology, West China Hospital of Sichuan University, Chengdu, China.
Department of Radiology, Chongqing General Hospital, Chongqing University, Chongqing, China.
J Eur Acad Dermatol Venereol. 2025 Jul 18. doi: 10.1111/jdv.20850.
Anti-MDA5+ dermatomyositis (DM) represents a heterogeneous group of patients as diverse individuals display different clinical characteristics, disease progression and complications development. The heterogeneity makes it difficult to classify interstitial lung disease in anti-MDA5 antibodies positive dermatomyositis (anti-MDA5+ DM-ILD).
To distinguish and characterize phenotypic subgroups for anti-MDA5+ DM-ILD patients.
From August 2014 to March 2022, 188 anti-MDA5+ DM-ILD patients were retrospectively enrolled. 21 HRCT-derived quantitative features were reduced to four principal components through principal component analysis. The missForest algorithm was employed for missing data imputation. Clusters were determined by partitioning around medoids. The classification and regression tree (CART) algorithm was utilized to distinguish between three clusters.
The Silhouette Coefficient and Dunn index indicated an optimal cluster number of 3. Patients in cluster 2 (high-risk cluster) have the highest RP-ILD incidence rate (84.9% vs. 18.1% vs. 17.1%, cluster 2 vs. 1 vs. 3, p < 0.001), extremely high early mortality (88.7% vs. 4.3% vs. 2.4%, cluster 2 vs. 1 vs. 3, p < 0.001) and moderate dermato-rheumatologic pattern. Patients in cluster 1 correspond to a pure dermato-rheumatologic pattern (dermato-rheumatologic cluster). Cluster 3 (low-risk cluster) was characterized by not obvious dermato-rheumatologic and the lowest RP-ILD incidence (17.1%) as well as the lowest early mortality (2.4%). The accuracy of the CART algorithm in differentiating clusters was 67.7% in the validation cohort with 56 patients. Principal component 1 was a key feature in the CART algorithm. Sensitivity analyses employing multiple clustering approaches confirmed the robustness of the three-cluster solution by partitioning around medoids.
Clustering analysis offers valuable insights into the heterogeneity and clinical implications of anti-MDA5+ DM-ILD. HRCT-derived quantitative features demonstrate significant value for early risk stratification in anti-MDA5+ DM-ILD.
抗MDA5阳性皮肌炎（DM）患者群体具有异质性，不同个体表现出不同的临床特征、疾病进展和并发症发生情况。这种异质性使得抗MDA5抗体阳性皮肌炎相关间质性肺病（抗MDA5+ DM-ILD）的分类变得困难。
区分并描述抗MDA5+ DM-ILD患者的表型亚组。
回顾性纳入2014年8月至2022年3月期间的188例抗MDA5+ DM-ILD患者。通过主成分分析将21项高分辨率计算机断层扫描（HRCT）衍生的定量特征简化为四个主成分。采用缺失森林算法进行缺失数据插补。通过围绕中心点划分法确定聚类。利用分类与回归树（CART）算法区分三个聚类。
轮廓系数和邓恩指数表明最佳聚类数为3。聚类2（高风险聚类）中的患者有最高的快速进展性间质性肺病（RP-ILD）发病率（84.9% 对比18.1% 对比17.1%，聚类2对比1对比3，p < 0.001）、极高的早期死亡率（88.7% 对比4.3% 对比2.4%，聚类2对比1对比3，p < 0.001）以及中度皮肤-风湿病模式。聚类1中的患者对应纯皮肤-风湿病模式（皮肤-风湿病聚类）。聚类3（低风险聚类）的特征是皮肤-风湿病不明显、RP-ILD发病率最低（17.1%）以及早期死亡率最低（2.4%）。在包含56例患者的验证队列中，CART算法区分聚类的准确率为67.7%。主成分1是CART算法中的关键特征。采用多种聚类方法的敏感性分析通过围绕中心点划分法证实了三聚类解决方案的稳健性。
聚类分析为抗MDA5+ DM-ILD的异质性和临床意义提供了有价值的见解。HRCT衍生的定量特征在抗MDA5+ DM-ILD的早期风险分层中显示出重要价值。
