A validated clinical prediction model for vaginal fungal positivity using routine fluorescence microecological indicators

Abstract Existing predictive models for vulvovaginal candidiasis are predominantly single-center and lack temporal or external validation. We aimed to develop and validate a clinical prediction model for vaginal fungal positivity using routine fluorescence-based vaginal microecological indicators. This retrospective multicenter cohort study included 19,333 specimens from a development cohort, an internal temporal validation cohort, and a single-site external validation cohort. The development cohort comprised 7,140 specimens collected in 2024, and the internal temporal validation cohort included 6,809 specimens from 2025. The external validation cohort consisted of 5,384 specimens from an independent tertiary hospital. Three predictive models were developed using six shared indicators, including logistic regression, LASSO logistic regression, and LightGBM. Performance was assessed by Area Under Curve (AUC), Brier score, calibration curves, and decision curve analysis. A nomogram was constructed to facilitate individual-level risk estimation, and SHAP analysis was applied for model interpretation. Fungal positivity rates were 24.6%, 24.0%, and 17.1% across the three cohorts. LASSO retained all six predictors. In internal temporal validation, all models achieved acceptable discrimination (LR/LASSO: AUC = 0.711, 95% CI: 0.698–0.726; LightGBM: 0.720, 95% CI: 0.706–0.734). In the single-site external validation cohort, discrimination was good (LR/LASSO: AUC = 0.821, 95% CI: 0.808–0.833; LightGBM: 0.813, 95% CI: 0.797–0.828). Brier scores ranged from 0.123 to 0.163; calibration-in-the-large was significantly negative in external validation (systematic overprediction), indicating a need for recalibration at sites with different prevalence. At the high-sensitivity threshold, the LR model achieved sensitivity of 0.946 and NPV of 0.977 externally. Decision curve analysis indicated positive estimated net benefit over reference strategies across specified threshold ranges in both validation cohorts. Vaginal cleanliness grade was the strongest predictor (OR = 3.491, 95% CI: 3.104–3.926), followed by age, lactobacillus abundance, and three markers of bacterial or protozoal co-infection. SHAP analysis confirmed directional consistency across all predictors. A six-variable prediction model demonstrated acceptable internal temporal discrimination and good discrimination in a single-site external validation cohort, although systematic overprediction was observed externally. The accompanying nomogram facilitates individual-level risk estimation without computational tools. The model warrants prospective validation and local recalibration before consideration as an aid for risk stratification and intensified microscopic review.