[Construction of a high-efficiency risk-prediction model for urinary tract infection after BPH surgery based on machine learning algorithm]

To explore the risk factors for urinary tract infection (UTI) after benign prostatic hyperplasia (BPH) surgery, and construct a high-efficiency risk-prediction model based on machine learning algorithms (MLA). Methods This retrospective study included 200 cases of BPH surgically treated in Xinyu Hospital of Traditional Chinese Medicine from May 2020 to April 2023. Twenty-three cases were included in the UTI group, and 177 were enrolled into the non-UTI group. The relevant data were collected and analyzed for the statistically significant factors by univariate logistic analysis by using the SPSS software, MLA-based logistic regression, back-propagation (BP) neural networks and decision classification and regression tree (CRT). And the predictive values of the models established via different algorithms using the area under the receiver operating characteristic curve (AUC) were compared. We assessed the prediction accuracy of the models and identified one with the best prediction efficiency based on the mean absolute error (MAE). Results Univariate analysis indicated statistically significant differences between the UTI and non-UTI groups in age, comorbid diabetes mellitus (DM), urinary catheter-indwelling time, prostate volume, preoperative catheterization, preoperative IPSS (P < 0.05). The independent predictive variables for UTI after BPH surgery were shown to be age, IPSS, comorbid DM and prostate volume by the method of multivariate logistic regression model (P<0.05). Age, urinary catheter-indwelling time, prostate volume and IPSS were assessed as the influence factors by the CRT model (P<0.05), and prostate volume, IPSS, age and urinary catheter-indwelling time were assessed as the influence factors by the BP neural network model (P<0.05). Among the 3 risk-prediction models, the one constructed with the BP neural networks exhibited the best prediction efficiency (AUC: 0.992, the optimal truncation value: 0.912, corresponding sensitivity and specificity: 0.957 and 0.955). Conclusion The risk-prediction model constructed by MLA and BP neural networks based on the characteristic factors of age, preoperative urinary retention catheterization, urinary catheter-indwelling time, IPSS and comorbid DM has a high predictive value for UTI after BPH surgery which can be applied to the identification and management of such high-risk population.