BACKGROUND: Surgical site infections are common postoperative complications that often develop after hospital discharge. Timely diagnosis and optimal treatment choice are crucial for clinical success and cost-effectiveness in surgical site infection treatment. Computer vision and artificial intelligence methods have demonstrated their effectiveness in analyzing chronic wounds, but their applicability to postoperative wound assessment remains poorly understood. AIM: To compare the diagnostic accuracy of various machine learning models for the classification of surgical wound images to determine the presence or absence of surgical site infections in trauma and orthopedic patients. METHODS: The study included patients aged ≥18 years who were hospitalized at two clinical centers for joint replacement, metal osteosynthesis, spine decompression-stabilization, or other interventions, or for the treatment of surgical site infections following these procedures. The following machine learning algorithms were used for infection recognition: support vector machines (SVMs), logistic regression (LR), random forests (RF), convolutional neural networks (VGG16 + CNN), and a model with an attention mechanism (YOLO 11s-cls). RESULTS: The study group included 183 patients with surgical site infections, and the control group consisted of 115 patients without surgical site infections. A total of 512 surgical wound images (292 with surgical site infection and 220 without infection) obtained from 298 patients were included in the study. After augmentation, the training set comprised approximately 2500 images. The YOLO 11s-cls model demonstrated the best metrics on the test set: sensitivity 91.2%, accuracy 91%; F1-score 89%. However, the differences between this model and the VGG16 + CNN neural network in terms of sensitivity, specificity, and accuracy were not significant. For the remainig models, sensitivity ranged from 69.6% (RF) to 87% (VGG16 + CNN), accuracy ranged from 68% (RF) to 85% (VGG16 + CNN), and F1-score ranged from 66% (RF) to 83% (VGG16 + CNN). CONCLUSION: The potential of using machine learning methods for remote monitoring of surgical wounds in trauma and orthopedic patients was confirmed. The developed models can be used to create a multimodal system for assessing and monitoring wound infection after surgical interventions.
DOI: 10.17816/DD693591
Download the article (PDF) or read online at the publishing house website: https://journals.eco-vector.com/DD/article/view/693591
Download the full issue (PDF) from the publishing house website: https://journals.eco-vector.com/DD/issue/viewIssue/14657/9960
Download the article (PDF) from eLibrary (in Russian, registration required): https://elibrary.ru/item.asp?id=89350203
Nazarenko A. G., Kleymenova E. B., Molodchenkov A. I., Gorbatyuk D. S., Enikeev A. D., Kislyakov V. A., Yashina L. P. Machine learning methods for recognizing surgical site infection in trauma and orthopedic patients: a cross-sectional study // Digital Diagnostics. — 2026. — Vol. 7. — N. 1. — P. 39–54. doi: 10.17816/DD693591