Development and Validation of a Multimodal Artificial Neural Network Model for the Diagnosis of Bladder Outlet Obstruction and Detrusor Underactivity in Men with Lower Urinary Tract Symptoms: Protocol for a Diagnostic Study.

BACKGROUND: Bladder outlet obstruction (BOO) and detrusor underactivity (DU) can produce overlapping clinical and uroflowmetric presentations in men with lower urinary tract symptoms (LUTS). Pressure-flow studies (PFS) provide reference diagnoses, but their invasiveness and cost constrain access. Building on a single-center model that used only structured clinical, uroflowmetric, and ultrasonographic variables, this protocol adds uroflowmetry waveforms and cystoscopic images within multimodal artificial neural networks (ANNs). Whether these inputs improve diagnostic estimation beyond structured-data-only models requires evaluation in men undergoing clinically indicated cystoscopy, a selected population distinct from the broader LUTS population for whom noninvasive screening is ultimately intended. OBJECTIVE: To develop and internally validate a multimodal ANN that estimates separate, non-mutually exclusive probabilities of BOO and DU; to compare its discrimination with structured-data-only and modality-ablation models; and to evaluate the version-locked model prospectively against PFS-based reference diagnoses. METHODS: This two-phase study was approved by the Persahabatan General Hospital Ethics Committee (0222/KEPK-RSUPP/07/2026). Models are developed retrospectively using records of adult men with LUTS evaluated at Persahabatan General Hospital, Indonesia, in 2024-2025, and evaluated prospectively in a separate cohort enrolled from 2026 at the development site and three additional tertiary hospitals. Predictors include the International Prostate Symptom Score, comorbidities, prostate volume, free-uroflowmetry curves, and cystoscopic images. Two models are developed, one for each diagnostic target, each using parallel neural encoders with intermediate feature fusion. Reference diagnoses are assigned from PFS recordings using prespecified bladder outlet obstruction index and bladder contractility index criteria, with blinded expert review and adjudication. Internal validation uses patient-level nested stratified five-fold cross-validation, with preprocessing and hyperparameter tuning confined to training partitions. Comparators include penalized logistic regression and modality-ablation models. After internal validation, the model architecture, hyperparameters, and decision thresholds will be locked before any prospective enrollment or exposure to prospective outcome data; the lock date and version will be reported, and no prospective data, including interim performance, will be used to modify the models, which will be applied to new participants without retraining. Discrimination, calibration, and decision-curve net benefit will be reported separately for each diagnostic target and study phase. RESULTS: Retrospective cohort assembly and data curation were completed in August 2026, 250 participants enrolled. Development and internal validation of the multimodal models with cystoscopy and uroflowmetry waveform are scheduled for completion in December 2026, after which the models and decision thresholds will be locked. Final external validation analyses are planned for mid-2027. No other protocol deviations have occurred. CONCLUSIONS: The proposed multimodal ANN may support pre-PFS differentiation of BOO and DU in selected men with LUTS and facilitate mechanism-based clinical decision-making.