Division of Rheumatology, Department of Internal Medicine, Tokyo Women's Medical University School of Medicine, Tokyo, Japan.
Institute of Rheumatology, Tokyo Women's Medical University Hospital, Tokyo, Japan.
Rheumatology (Oxford). 2023 Jun 1;62(6):2272-2283. doi: 10.1093/rheumatology/keac586.
To construct a predictive model for the Sharp/van der Heijde score (SHS) and assess its applicability in clinical research settings.
A prediction model for SHS was constructed in three steps using convolutional neural networks (CNN) and an in-house RA image database: orientation, detection and damage prediction. A predictive model for radiographic progression (ΔSHS >3/year) was developed using a graph convolutional network (GCN). A multiple regression model was used to assess the association between predicted SHS using the CNN model and clinical features.
In the orientation and detection phases, 100% accuracy was achieved in the image orientation correction, and all predicted joint coordinates were within 10 pixels of the correct coordinates. In the damage prediction phase, the κ values between the model and expert 1 were 0.879 and 0.865 for erosion and joint space narrowing, respectively. Using a dataset scored by experts 1 and 2, a minimal overfitting was determined to the scoring by expert 1. High-titre RF was an independent risk factor of ΔSHS per year, as predicted by the CNN model in biologics users. The AUCs of the GCN model for predicting ΔSHS >3/year in patients with and without biologics at baseline were 0.753 and 0.734, respectively, superior to those of the other models. The RF titre was the most important feature in predicting ΔSHS >3/year in biologics users in the GCN model.
A high-performance scoring model for SHS that is applicable to clinical research was constructed.
构建 Sharp/van der Heijde 评分（SHS）的预测模型，并评估其在临床研究中的适用性。
使用卷积神经网络（CNN）和内部 RA 图像数据库，分三个步骤构建 SHS 的预测模型：方向、检测和损伤预测。使用图卷积网络（GCN）开发用于预测放射学进展（ΔSHS>3/年）的预测模型。使用多元回归模型评估使用 CNN 模型预测的 SHS 与临床特征之间的关联。
在方向和检测阶段，图像定向校正的准确率达到 100%，所有预测的关节坐标均在正确坐标的 10 个像素以内。在损伤预测阶段，模型与专家 1 之间的κ值分别为 0.879 和 0.865，用于侵蚀和关节间隙狭窄。使用由专家 1 和 2 评分的数据集，确定对专家 1 评分的最小过度拟合。高滴度 RF 是生物制剂使用者中 CNN 模型预测的每年 ΔSHS 的独立危险因素。在基线时使用和不使用生物制剂的患者中，GCN 模型预测ΔSHS>3/年的 AUC 分别为 0.753 和 0.734，优于其他模型。在 GCN 模型中，RF 滴度是预测生物制剂使用者中ΔSHS>3/年的最重要特征。
构建了适用于临床研究的高性能 SHS 评分模型。
Zotero 插件
