Lim Jongmin, Chang Sehun, Kim Kwangsu, Park Hee Jin, Kim Eugene, Hong Seok Woo
Department of Computer Science and Engineering, Sungkyunkwan University College of Computing and Informatics, Suwon, South Korea.
Department of Radiology, Kangbuk Samsung Hospital, Sungkyunkwan University School of Medicine, Seoul, South Korea.
J Orthop Surg Res. 2025 Apr 26;20(1):419. doi: 10.1186/s13018-025-05830-z.
Treatments for distal radius fractures (DRFs) are determined by various factors. Therefore, quantitative or qualitative tools have been introduced to assist in deciding the treatment approach. This study aimed to develop a machine learning (ML) model that determines the need for surgical treatment in patients with DRFs using a ML model that incorporates various clinical data concatenated with plain radiographs in the anteroposterior and lateral views.
Radiographic and clinical data from 1,139 patients were collected and used to train the ML models. To analyze and integrate data effectively, the proposed ML model was mainly composed of a U-Net-based image feature extractor for radiographs, a multilayer perceptron based clinical feature extractor for clinical data, and a final classifier that combined the extracted features to predict the necessity of surgical treatment. To promote interpretability and support clinical adoption, Gradient-weighted Class Activation Mapping (Grad-CAM) was employed to provide visual insights into the radiographic data. SHapley Additive exPlanations (SHAP) were utilized to elucidate the contributions of each clinical feature to the predictions of the model.
The model integrating image and clinical data achieved accuracy, sensitivity, and specificity of 92.98%, 93.28%, and 92.55%, respectively, in predicting the need for surgical treatment in patients with DRFs. These findings demonstrate the enhanced performance of the integrated model compared to the image-only model. In the Grad-CAM heatmaps, key regions such as the radiocarpal joint, volar, and dorsal cortex of the radial metaphysis were highlighted, indicating critical areas for model training. The SHAP results indicated that being female and having subsequent or concomitant fractures were strongly associated with the need for surgical treatment.
The proposed ML models may assist in assessing the need for surgical treatment in patients with DRFs. By improving the accuracy of treatment decisions, this model may enhance the success rate of fracture treatments, guiding clinical decisions and improving efficiency in clinical settings.
桡骨远端骨折（DRF）的治疗由多种因素决定。因此，已引入定量或定性工具来辅助确定治疗方法。本研究旨在开发一种机器学习（ML）模型，该模型使用结合了前后位和侧位平片的各种临床数据的ML模型来确定DRF患者是否需要手术治疗。
收集了1139例患者的影像学和临床数据，并用于训练ML模型。为了有效地分析和整合数据，所提出的ML模型主要由用于X线片的基于U-Net的图像特征提取器、用于临床数据的基于多层感知器的临床特征提取器以及结合提取的特征以预测手术治疗必要性的最终分类器组成。为了提高可解释性并支持临床应用，采用梯度加权类激活映射（Grad-CAM）对影像学数据提供可视化见解。利用SHapley加性解释（SHAP）来阐明每个临床特征对模型预测的贡献。
在预测DRF患者的手术治疗需求时，整合图像和临床数据的模型的准确率、敏感性和特异性分别达到92.98%、93.28%和92.55%。这些结果表明，与仅使用图像的模型相比，整合模型的性能有所提高。在Grad-CAM热图中，桡腕关节、桡骨干骺端掌侧和背侧皮质等关键区域被突出显示，表明这些是模型训练的关键区域。SHAP结果表明，女性以及有后续或合并骨折与手术治疗需求密切相关。
所提出的ML模型可能有助于评估DRF患者的手术治疗需求。通过提高治疗决策的准确性，该模型可能提高骨折治疗的成功率，指导临床决策并提高临床环境中的效率。
