Non-invasive parameters for the prediction of urodynamic bladder outlet obstruction: analysis using causal Bayesian networks

Abstract

Purpose: Numerous attempts have been made to predict urodynamic bladder outlet obstruction (BOO), however, little information exists on non-invasive parameters for BOO prediction. We aimed to identify non-invasive clinical parameters to predict BOO using causal Bayesian networks (CBN).

Methods: From October 2004 to December 2011, patients with lower urinary tract symptoms (LUTS) suggestive of BPH were included in this study. Out of the 1352 patients, 866 were selected for the analysis. Mean age, total prostate volume (TPV) and IPSS were 66.3 (±7.0, SD) years, 49.8 (±26.7) ml, and 18.0 (±7.7), respectively. Mean bladder outlet obstruction index (BOOI) was 34.0 (± 24.4), and 292 (33.5%) patients had urodynamic BOO (BOOI ≥40). Non-invasive predictors of BOO were selected using CBN. BOO prediction with selected parameters was verified using logistic regression (LR) and artificial neural networks (ANN) considering whole non-invasive parameters.

Results: CBN identified TPV, Qmax, PVR, and IPSS item 5 (slow-stream) as independent predictors of BOO. With these four parameters, sensitivity and specificity of BOO prediction were 54.1% and 86.4%, respectively, with an area under receiver operating characteristic curve (AUROC) of 0.793. LR and ANN models with the same parameters showed similar accuracy (LR: sensitivity 51.7%, specificity 90.9%, AUROC 0.797

ANN: sensitivity 43.7%, specificity 92.7%, AUROC 0.756). The AUROC of ANN was smaller than that of the other two methods (p-value range <0.001-0.005).

Conclusions: Our study demonstrated that TPV, Qmax, PVR, and IPSS item 5 (slow-stream) are independent predictors of urodynamic BOO.