Construction of interpretable machine-learning diagnostic models for erectile dysfunction based on routine blood and biochemical detection data

Despite its high incidence, the diagnosis of erectile dysfunction (ED) is impeded by current diagnostic constraints and patient hesitancy in seeking medical intervention. The integration of machine learning with standard blood and biochemical markers has demonstrated diagnostic efficacy across a spectrum of diseases, yet its application in ED remains unexplored. Based on data from the 2001-2004 National Health and Nutrition Examination Survey, male respondents who did not meet the study criteria were excluded. The data set was allocated to training and test subsets at a proportion of 7 to 3. The Boruta algorithm was used for feature selection. Stepwise regression was used to initially screen for significant variables. Iterative testing and elimination based on the VIF and random forest, logistic regression, XGBoost, SVM, Gradient Boost, LightGBM, and CatBoost models were used to construct the models and evaluate their performance. The models were further interpreted using SHapley Additive exPlanations analysis. This study included 945 individuals with ED and 2520 individuals without ED. Most of the routine blood and biochemical parameters differed between the non-ED and ED groups. The standardized net benefit thresholds for applying the CatBoost model ranged from 0 to 75%. Analysis of routinely measured blood and biochemical parameters revealed correlations of elevated glycohemoglobin, lactate dehydrogenase, and erythrocyte distribution width levels with increased susceptibility to ED. This study reveals the potential role of routine blood and biochemical markers in the etiology of ED.