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

Comprehensive Summary

This study developed interpretable machine learning models to predict erectile dysfunction (ED) using routine blood and biochemical data. The authors used NHANES data, including 945 men with ED and 2520 without, and applied multiple machine learning methods such as random forest, logistic regression, XGBoost, SVM, and CatBoost. Feature selection was done using the Boruta algorithm, and model outputs were interpreted using SHAP analysis. Among the models tested, CatBoost showed the best overall performance, with consistent predictive value across a wide range of thresholds. The analysis identified several key factors associated with ED, including higher levels of glycohemoglobin, lactate dehydrogenase, and erythrocyte distribution width. These findings suggest that commonly available lab markers, which are not traditionally used to diagnose ED, may still carry useful diagnostic information. Overall, the study shows that combining routine clinical data with machine learning can help detect patterns linked to ED risk.

Outcomes and Implications

ED diagnosis is often limited by underreporting and reliance on subjective assessments, since many patients avoid seeking care. This approach shifts the focus toward objective, routinely collected lab data, which could allow earlier identification of at-risk patients without relying on self-report.The use of interpretable models (via SHAP) is also important. Instead of just predicting risk, the model shows which variables are driving the prediction, making it easier to understand how metabolic and systemic health factors relate to ED. This fits into a broader trend where machine learning is being used not just for prediction, but for uncovering relationships that standard statistical models might miss. That said, this is still based on retrospective population data, not real-time clinical deployment. So while the model is promising, it’s not something you’d plug into clinic workflow yet without further validation.