Prospective multicenter development and external validation of a narrow-band imaging based prediction model for malignancy risk in bladder lesions.

Abstract

644 Background: White-light cystoscopy lacks specificity, driving unnecessary resections. NBI improves visibility but raises false positives and lacks a thresholded aid. We externally validated an NBI rule-in nomogram for real-time use. Methods: Prospective, seven-center, lesion-level study (11/2021–06/2024). Consecutive adults undergoing cystoscopy were scored in vivo under NBI with a standardized atlas for morphology (smooth/cauliflower/patchy) and vascular patterns (dotted/thick-branching); lesion size and routine covariates were recorded on harmonized CRFs; uniform Olympus CYF-VHA across sites. Histopathology after resection/biopsy was the reference. A prespecified logistic nomogram was trained and validated with centers held out. Analyses (R v4.2) included AUC (DeLong), calibration (intercept/slope/Brier/HL), decision-curve per-100 translation, and ≤10/≤5-mm subgroups. Results: Seven-center study (n = 1,404 lesions; 843 patients) showed a prespecified 7-feature NBI nomogram, at fixed threshold 0.537, with external performance specificity 0.918 and sensitivity 0.861. Multivariable effects (OR, 95% CI) aligned with endoscopic heuristics: cauliflower-like 10.83 (4.79–26.00); dotted 6.77 (3.27–14.67); thick-branching 4.14 (2.17–8.18); gross hematuria 3.19 (1.55–6.77); boundary-clear 2.17 (1.11–4.32); male 2.16 (1.09–4.33); patchy 0.13 (0.02–0.53). Diagnostic transforms supported rule-in use (LR⁺ 10.47, Youden’s J 0.779). Per 100 cystoscopies, ~38.8 benign resections avoided and ~8 cancers missed; training/internal showed similar magnitudes (~32.4/~66.7 avoided; ~6.4/~5.3 missed), concordant with net benefit at the same threshold (external ~31.6 fewer unnecessary interventions per 100). Discrimination was high: AUC 0.93 (95% CI 0.91–0.95) training, 0.98 (0.96–0.99) internal, 0.95 (0.93–0.97) external. Training/internal threshold performance is consistent across cohorts (sensitivity 0.827–0.896, specificity 0.841–0.958; LR⁺ 5.63–19.68) and centers. Calibration: validation Brier 0.06–0.10, Hosmer–Lemeshow p > 0.05 across cohorts, supporting the cut-point. Performance was preserved in challenging strata: ≤10 mm lesions AUC 0.927 (0.906–0.948) with sensitivity/specificity 0.846/0.888; ≤5 mm AUC 0.925 (0.892–0.957) with 0.878/0.893. EPV≈60 mitigated overfitting; all inputs are immediately observable under NBI. Conclusions: An externally validated, multicenter NBI-based prediction model converts familiar endoscopic cues into a prespecified, scope-ready rule-in decision that curbs unnecessary interventions while maintaining clinical safety, even in sub-centimeter lesions. Inputs are directly observable, computation is immediate, and calibration/validation support transportable use across operators and sites; a prospective implementation study is warranted.