In male lower urinary tract symptoms/benign prostatic hyperplasia (LUTS/BPH), de-obstructive surgery can predict a clear improvement in voiding symptoms when detrusor contractility is normal and only bladder outlet obstruction is present. However, it is difficult to accurately predict improvement in voiding symptoms when detrusor contractility is impaired. In this study, we aimed to develop an artificial intelligence model to predict the improvement (degree) of symptoms after holmium laser enucleation of the prostate (HoLEP) based on the degree of bladder contractility and to assess the prediction of change in maximum flow rate and voiding efficiency at 1month after surgery.

We screened 1933 patients analyses performed in Samsung Medical Center from July 2008 to January 2024. The deep neural network for multi-class classification to predict simultaneously both amount of change in the maximum flow rate and voiding efficiency which were respectively classified into three classes was employed to predict the recovery rate in detail. Furthermore, the machine learning algorithms that can be applied to multi-class classification as Extreme Gradient Boosting, Random Forest Classification and Support Vector Machine were applied for comparative analysis. In order to mitigate over-fitting resulting from class imbalance, we proposed that the least squares method and multi task learning for the deep neural network were employed to address this challenge.

We included 1142 patients without missing data and separated 992 patients for model training from July 2008 to December 2022 and 150 patients for external validation from January 2023 to January 2024. The deep neural network for multi-class classification was obtained a micro-AUC of 0.890 ± 0.014 (0.867), micro-sensitivity of 0.799 ± 0.032 (0.700) and micro-specificity of 0.899 ± 0.016 (0.850) for amount of change in the maximum flow rate and an micro-AUC of 0.830 ± 0.005 (0.695), micro-sensitivity of 0.670 ± 0.016 (0.580) and micro-specificity of 0.835 ± 0.008 (0.790) for voiding efficiency.

The deep neural network for multi-class classification supports the prediction in detail. The least square method prevent while over-fitting occur by class imbalance. This model can provide prediction guidelines to primary caregivers and assist in diagnosis of lower urinary tract symptoms /benign prostatic hyperplasia after the operation.