Artificial inteligence reading of cystometric traces provides good correlation with human diagnosis

Urodynamic studies are essential for diagnosing lower urinary tract dysfunction but are expert-dependent and time-consuming. Artificial intelligence (AI), notably machine learning (ML) and deep learning (DL) may help automate and standardize interpretation, reducing inter-observer variability and improving efficiency.

To evaluate the correlation between artificial intelligence (AI) based classification and human expert diagnosis of detrusor overactivity (DO) in cystometry (CMG), with explicit handling of artifacts and quantification of parameters.

Retrospective, single-center, observational diagnostic-accuracy (cross-sectional) study with a consecutive cohort of adults who underwent cystometry in 2023, in which AI outputs were compared with a reference standard (three-urologist consensus). We evaluated 517 cystometry (CMG) tracings: 200 used to train AI models and 317 reserved for testing. Two approaches were assessed: (i) image-based CNN-VGG16 deep learning, which achieved 75% accuracy for detecting detrusor overactivity (DO) but did not yield quantitative metrics and (ii) wavelet-based ML (Daubechies transforms), which improved accuracy to 84.2%, with 82.6% specificity and 86.3% sensitivity, while providing detailed contraction descriptors. An Isolation Forest anomaly-detection stage identified and managed artifacts (e.g., coughs, open lines, catheter movement). Integrating signal processing (time-frequency denoising and rule-based thresholds) with AI classification supported robust CMG event recognition, enabling clearer identification of DO, estimation of bladder compliance from DO-free segments, and mitigation of artifacts. Both branches produced classifications in less than 20 s per study.

Combining algorithmic outputs with expert supervision could deliver practical, faster, and more reproducible urodynamic reporting, while preserving clinical accountability and transparency and generalizability.