Identification of factors associated with failure to meet nocturnal penile tumescence and rigidity criteria based on RigiScan monitoring and machine learning algorithms: a cross-sectional study

Background Using RigiScan nocturnal penile tumescence and rigidity (NPTR) monitoring as reference, we applied machine learning to identify routine clinical indicators associated with failure to meet NPTR criteria (no effective nocturnal erection). Methods A total of 519 patients with erectile dysfunction completed two consecutive nights of RigiScan monitoring and were divided into effective and no-effective nocturnal erection groups. Data were split 7:3 into training (n = 364) and internal validation (n = 155) sets. Feature selection (univariate logistic regression, Boruta, LASSO) was performed exclusively in the training set. Five models (LR, RF, SVM, LightGBM, XGBoost) were constructed. The classification threshold was determined in the training set and fixed in the validation set. Performance was assessed using AUC with bootstrap 95% CIs, Brier score, calibration slope and intercept, sensitivity, specificity, PPV, and NPV. Nested cross-validation (5 × 3) was performed. SHAP interpreted the best model. Results The effective group comprised 389 patients and the no-effective group 130. Six variables were selected: HbA1c, GLU, Smoke, Hypertension, Age, and BMI. In the validation set, LR achieved the highest AUC (0.779, 95% CI 0.692–0.861), followed by LGBM (0.718), XGB (0.700), RF (0.699), and SVC (0.646). DeLong’s test showed LR significantly outperformed RF (p = 0.028) and SVC (p = 0.009), but not LGBM or XGB. In nested cross-validation, LR also had a higher mean AUC than RF (0.726 ± 0.052 vs 0.666 ± 0.034). LR showed the lowest Brier score (0.151) and calibration slope closest to 1 (1.258). SHAP identified BMI, GLU, and Age as the top three contributors. Conclusions Conventional metabolic indicators achieved only moderate discrimination. BMI, GLU, and HbA1c were consistently associated with failure to meet NPTR criteria. LR was the most robust model but cannot replace RigiScan monitoring. These findings should be interpreted as associations, not causal relationships.