A pilot study: Incorporating Treponema pallidum antigens into machine learning models for accurate syphilis treatment outcome assessment.

As the reliability of nontreponemal tests for evaluating syphilis treatment efficacy is increasingly questioned, we propose an optimized approach using Treponema pallidum (Tp) antigens (Tp0134, Tp0768, Tp0971, Tp0462, and Tp92) combined with a machine learning (ML) model. Analysis of 509 serum samples (including paired pre- and post-treatment samples) employed an established ELISA assay to dynamically monitor antibody changes. Results demonstrated that post-treatment antibody reduction for potential infection-stage-dependent antigens (pIDAs) (especially Tp0134 and Tp0768) was markedly higher than for the non-infection-stage-dependent antigen Tp92 and traditional methods. Utilizing nested cross-validation to train an array of ML models, ultimately chosen random forest model (AUC = 0.815) demonstrated enhanced efficacy in accurately distinguishing between infection and cure. Specifically, Tp0768, Tp92, and Tp0134 were identified as the pivotal features. Combining Tp antigens with ML provides a more accurate and dynamic tool for treatment efficacy assessment, enabling a more effective evaluation of syphilis treatment outcomes in the future.