Optimizing machine learning-based diagnosis of neurosyphilis in HIV-negative patients: a multicenter, real-world comparison of international diagnostic criteria

Clinical Snapshot

Key Findings

• P Value: Not reported in abstract

• Effect Size: AUC and PRAUC >0.90 for all models

• Primary Outcome: Diagnostic accuracy across six international guidelines using machine learning models

• Nnt Or Sensitivity: Sensitivity and specificity values not specified in abstract, though excellent discrimination (AUC >0.90) suggests high diagnostic accuracy

• Confidence Interval: Not reported in abstract

Clinical Application

High feasibility with free web-based tool requiring only basic clinical and CSF parameters Highly relevant for Australian practice given rising neurosyphilis rates; tool could support rural and remote clinicians with limited access to specialist expertise; aligns with RACGP emphasis on evidence-based diagnostic support tools HIV-negative patients with suspected neurosyphilis in both resource-rich and resource-limited healthcare settings

Abstract

Neurosyphilis continues to rise globally, yet diagnosis remains challenging, often requiring multidisciplinary expertise and multiple CSF assays that are difficult to access in resource-limited settings. We developed and compared machine-learning (ML) models tailored to six international diagnostic guidelines and built a free, web-based tool for guideline-adapted decision support. We assembled 1,648 suspected neurosyphilis cases from four centers, using Guangzhou as the training cohort and Beijing, Xiamen (China) and Seattle (USA) as external validation cohorts. Five algorithms (Random Forest, AdaBoost, SVC, NuSVC, XGBoost) were trained with randomized search and three-fold cross-validation; performance was assessed by AUC, PRAUC, calibration, decision-curve net benefit, Brier score, and SHAP for model explainability. Across models and guidelines, neurological symptoms, CSF protein, and CSF white blood cell count consistently ranked as the strongest predictors. All models achieved excellent discrimination (AUC and PRAUC >0.90) with good calibration, reliability, and positive clinical utility, though performance varied modestly by guideline. These findings indicate that the optimal ML approach depends on the diagnostic definition applied. Our freely available online tool operationalizes these models to provide clinicians worldwide with context-adapted support aligned to local criteria.