Using machine learning to construct the diagnosis model of female bladder outlet obstruction based on urodynamic study data

Abstract

Purpose: To intelligently diagnose whether there is bladder outlet obstruction (BOO) in female with decent detrusor contraction ability by focusing on urodynamic study (UDS) data. Materials and Methods: We retrospectively reviewed the UDS data of female patients during urination. Eleven easily accessible urinary flow indicators were calculated according to the UDS data of each patient during voiding period. Eight diagnosis models based on back propagation neural network with different input feature combination were constructed by analyzing the correlations between indicators and lower urinary tract dysfunction labels. Subsequently, the stability of diagnostic models was evaluated by five-fold cross-validation based on training data, while the performance was compared on test dataset. Results: UDS data from 134 female patients with a median age of 51 years (range, 27–78 years) were selected for our study. Among them, 66 patients suffered BOO and the remaining were normal. Applying the 5-fold cross-validation method, the model with the best performance achieved an area under the receiver operating characteristic curve (AUC) value of 0.949±0.060 using 9 UDS input features. The accuracy, sensitivity, and specificity for BOO diagnosis model in the testing process are 94.4%, 100%, and 89.3%, respectively. Conclusions: The 9 significant indicators in UDS were employed to construct a diagnostic model of female BOO based on machine learning algorithm, which performs preferable classification accuracy and stability.

Keywords

• back propagation neural network
• bladder outlet obstruction
• feature selection
• machine learning
• urodynamic study