CONSTRUCTION OF A SYPHILIS RISK CALCULATION MODEL AND DERIVATION OF KNOWLEDGE RULES

Objective: This study aims to construct a computational model using machine learning, uncover syphilis risk factors, analyze and predict syphilis risk, and propose corresponding measures for syphilis prevention and intervention. Method: Retrospective data from 2000 valid cases, each containing 16 indicators, were collected through Voluntary Counseling and Testing (VCT) for syphilis. Logistic regression analyses were initially used to identify statistically significant risk factors for syphilis infection ([Formula: see text]) and to analyze sources of risk. Subsequently, decision trees and random forest models were established using logistic single-factor regression analysis with relevant variables as independent variables and syphilis infection as the dependent variable. Additionally, systematic approaches targeting the issue of imbalanced syphilis data were undertaken, including sampling, encoding, and algorithmic adjustments. Results: Univariate logistic regression analysis showed that age, occupation, age of first sexual encounter, sexual orientation, attendance at HIV-related lectures, awareness of condoms, awareness of lubricants, and history of sexually transmitted infections were statistically associated with syphilis infection ([Formula: see text]). Multivariate logistic regression identified only age and sexual orientation as independent predictors significantly associated with syphilis infection. Model performance evaluation based on ROC curves, confusion matrices, and derived metrics indicated that the random forest model outperformed the decision tree. Furthermore, the five high-risk classification rules generated by the selected decision tree model largely corresponded to the risk factors identified in the univariate logistic regression analysis. Conclusion: From the logistic regression analysis, significant risk factors associated with syphilis were accurately identified, effectively achieving dimensionality reduction and feature selection, thereby laying a solid foundation for constructing high-performance machine learning classification models. In terms of classification performance, the random forest model demonstrated superior predictive accuracy, while the decision tree generated interpretable classification rules that enhanced model transparency. Together, these approaches enabled a comprehensive risk assessment, providing valuable insights and practical references for syphilis prevention strategies and research on syphilis infection risk management in China.