Abstract
BACKGROUND: Machine learning (ML) shows promise for sexually transmitted infection (STI) risk prediction, but systematic evidence of its effectiveness remains fragmented. METHODS: We systematically searched six electronic databases, three preprint archives and conference proceedings (January 2010-April 2024). Studies reporting quantitative performance metrics for supervised ML-based STI risk prediction models were included. We used a bivariate random-effects model to estimate pooled sensitivity, specificity and area under the curve (AUC). The risk of bias was assessed using the Prediction model Risk of Bias Assessment Tool. We conducted sequential analyses of studies with complete and reconstructed confusion matrices. Subgroup analyses and meta-regression explored potential sources of heterogeneity. RESULTS: Among 3877 records screened, 25 studies comprising 45 unique models met inclusion criteria. For HIV, analysis of studies with complete confusion matrices (7 studies, 9 contingency tables) demonstrated summary AUC of 0.91 (95% CI: 0.88-0.93), pooled sensitivity 0.84 (0.76-0.90) and specificity 0.84 (0.70-0.93). Substantial heterogeneity persisted across subgroups (I² > 98%). For other STIs, individual studies reported AUCs ranging from 0.75-0.87 for syphilis (n = 5), 0.73-1.00 for gonorrhoea (n = 6) and 0.67-1.00 for chlamydia (n = 6). DISCUSSION: While ML models show promising performance, particularly for HIV, significant heterogeneity complicates interpretation. Future research should prioritize external validation, standardized guidelines and multi-centred robust implementation studies to evaluate clinical impact.