(767) Comparing Machine Learning Models for Rule-Out Prediction of Sexually Transmitted Infections in Male Patients: A Retrospective Study Using Urinalysis and Symptom Data

Introduction Nucleic acid amplification tests (NAATs) are the gold standard for diagnosing sexually transmitted infections (STIs); however, their overuse in low-risk patients results in unnecessary costs and adds to laboratory workload. In clinical practice, NAATs are often ordered for men who present after suspected sexual exposure, including asymptomatic individuals and those with urethritis, urethral discharge, urethral tingling, or other genitourinary symptoms, even though the likelihood of infection varies widely. Developing a reliable tool to identify low-risk patients who can safely delay testing could improve diagnostic stewardship and cost-effectiveness. Objective Develop and temporally validate a machine-learning (ML)–based rule-out model that predicts a low likelihood of NAAT-detectable STIs using routine urinalysis and clinical symptom data. Methods A retrospective analysis was conducted on 532 male patients who underwent multiplex PCR–based NAAT for suspected STIs from January 2020 to May 2025. Clinical symptoms and urinalysis parameters-such as leukocyte esterase (LE), nitrite, urinary WBC, and RBC-were extracted from electronic records. Five supervised ML algorithms-logistic regression, random forest, XGBoost, support vector machine, and naïve Bayes-were trained on data from 2020–2024 and validated on data from 2025. Probability thresholds were calibrated to ensure ≥95% sensitivity and ≥90% negative predictive value (NPV). Model interpretability was assessed using SHAP (Shapley Additive Explanations). Results Among 532 patients, 34.8% had positive NAAT results. The most common detections were Ureaplasma spp. (50.4%), Mycoplasma hominis (15.0%), Mycoplasma genitalium (10.7%), and Chlamydia trachomatis (11.1%). LE and WBC positivity were significantly associated with NAAT positivity (P<0.001). The XGBoost model showed the best rule-out performance (AUC = 0.86; sensitivity = 0.98; NPV = 0.91), safely deferring 44.2% of NAAT tests. Urinary WBC, LE, and nitrite were the strongest predictors of infection, while asymptomatic presentation and lack of inflammation supported safe test deferral. SHAP interpretation confirmed the biological plausibility of these findings. Conclusions An interpretable ML model that combines urinalysis and symptom data accurately identifies male patients with low STI risk, reducing unnecessary NAAT testing by 44% while maintaining high sensitivity and NPV. This rule-out approach supports diagnostic stewardship and can be integrated into clinical workflows to improve cost-effective STI management. Disclosure No