Machine learning methods for recognizing surgical site infection in trauma and orthopedic patients

BACKGROUND: Surgical site infections (SSIs) 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 SSI treatment. Computer vision and artificial intelligence (AI) methods have demonstrated their effectiveness in analyzing chronic wounds, but their applicability to postoperative wound assessment remains poorly understood. AIM: A comparative analysis of performance metrics for various machine learning methods in classifying surgical wounds images for the SSI presence or absence in trauma and orthopedic patients. METHODS: The sample included 512 surgical wound images (292 with SSI and 220 without infection) obtained from 298 patients aged 18 years and older admitted to the N.N. Priorov National Medical Research Center of Traumatology and Orthopedics or the Purulent Surgery Department of the A.K. Eramishantsev City Clinical Hospital. After applying augmentation methods, approximately 2,500 images were obtained. The following machine learning algorithms were used for infection recognition: support vector machines (SVM), linear regression (LR), random forests (RF), convolutional neural networks (VGG16+CNN), and a model with an attention mechanism (YOLO 11s-cls). RESULTS: The YOLO 11s-cls model demonstrated the best metrics on the test set: sensitivity 91.2%, accuracy 91%; F1-score 89%; balanced accuracy 90.6%. For the other models, sensitivity ranged from 69.6% (RF) to 87% (VGG16+CNN), accuracy 68% (RF) to 85% (VGG16+CNN), and F1-score from 66% (RF) to 83% (VGG16+CNN). CONCLUSION: The study results confirmed the potential of AI appropriateness for remote monitoring of surgical wounds in trauma and orthopedic patients. The developed models can be used to create a multimodal system for assessing and monitoring wound infection after surgical interventions.