Establishment and application of TSDPSO-SVM model combined with multi-dimensional feature fusion method in the identification of fracture-related infection

Abstract

AbstractFracture-related infection (FRI) is one of the most common and intractable complications in orthopedic trauma surgery. This complication can impose severe psychological burdens and socio-economic impacts on patients. Although the definition of FRI has been proposed recently by an expert group, the diagnostic criteria for FRI are not yet standardized. A total of 4761 FRI patients and 4761 fracture patients (Non-FRI) were included in the study. The feature set of patients included imaging characteristics, demographic information, clinical symptoms, microbiological findings, and serum inflammatory markers, which were reduced by the Principal Component Analysis. To optimize the Support Vector Machine (SVM) model, the Traction Switching Delay Particle Swarm Optimization (TSDPSO) algorithm, a recognition method was proposed. Moreover, five machine learning models, including TSDPSO-SVM, were employed to distinguish FRI from Non-FRI. The Area under the Curve of TSDPSO-SVM was 0.91, at least 5% higher than that of other models. Compared with the Random Forest, Backpropagation Neural Network (BP), SVM and eXtreme Gradient Boosting (XGBoost), TSDPSO-SVM demonstrated remarkable accuracy in the test set ($$\chi^{ 2} = 29.17, \,50.46, \,56.66,\,35.88, P < 0.01$$ χ 2 = 29.17 , 50.46 , 56.66 , 35.88 , P < 0.01 ). The recall of TSDPSO-SVM was 98.32%, indicating a significant improvement ($$\chi^{2} = 91.78,\, 107.42, \,135.69, P < 0.01$$ χ 2 = 91.78 , 107.42 , 135.69 , P < 0.01 ). Compared with BP and SVM, TSDPSO-SVM exhibited significantly superior specificity, false positive rate and precision ($$\chi^{2} > 3.84, P < 0.05)$$ χ 2 > 3.84 , P < 0.05 ) . The five models yielded consistent results in the training and testing of FRI patients across different age groups. TSDPSO-SVM is validated to have the maximum overall prediction ability and can effectively distinguish between FRI and Non-FRI. For the early diagnosis of FRI, TSDPSO-SVM may provide a reference basis for clinicians, especially those with insufficient experience. These results also lay a foundation for the intelligent diagnosis of FRI. Furthermore, these findings exhibit the application potential of this model in the diagnosis and classification of other diseases.