预测西班牙裔男性前列腺癌风险的机器学习模型

Introduction: Hispanic men in the United States experience a notable prostate cancer (PCa) burden, with studies reporting higher rates of advanced-stage disease at diagnosis compared with non-Hispanic men. However, their genetic and sociodemographic profiles remain untargeted in current risk prediction approaches. These observations highlight gaps in current risk assessment strategies and the need for improved, population-relevant prediction tools. Machine learning (ML) applications in biomedical research have yielded valuable advances in the early detection and prediction of cancer risk. The aim of this study was to develop a ML model to predict PCa risk in Hispanic men leveraging data from the All of Us Research Program. Methods: The study cohort included 20,172 Hispanic males, 220 PCa cases (1.1%) and 19,952 controls. Predictors included age group (≤40, 41-59, ≥60 years), family history of PCa, smoking status, drinking frequency and the rs10090154 genotype associated with PCa. Data were split 80/20 into training (n= 16, 138) and testing (n= 4,034) sets. To address class imbalance, class weights were applied. Logistic regression, random forest and XGBoost models were trained and evaluated. Models underwent a five-fold cross-validation with sigmoid calibration, and performance was assessed using ROC-AUC, PR-AUC, Brier score, precision recall and F1 score. Results: Among the three models, the logistic regression model achieved the best balance of discrimination and calibration with a mean ROC-AUC of 0.872 (95% CI: 0.856-0.885), PR-AUC of 0.059 (95% CI: 0.045-0.074), Brier score of 0.0105 and an F1 score of 0.066. The model metrics were accuracy = 0.704, recall = 0.955 and precision = 0.034. Correspondingly, the false negative rate (FNR) was 4.5%, indicating that only 2 out of 44 true PCa cases were misclassified as controls, while the false positive rate (FPR) was 30.6%, meaning roughly one in three non PCa cases were incorrectly flagged as positive. Feature importance and odds ratio analyses identified Age ≥60 and family history of PCa as the most influential predictors for the model, followed by the rs10090154 genotype and former smoking status. Comparatively, the random forest (ROC-AUC= 0.851) and XGBoost (ROC- AUC= 0.858) models demonstrated similar discrimination but exhibited slightly lower calibration and interpretability compared with logistic regression. Conclusions: A class-weighted, sigmoid-calibrated logistic regression model achieved accurate, interpretable and well calibrated PCa risk predictions. Despite the low observed case prevalence in the All of Us cohort, the model showed strong sensitivity and discrimination, supporting its utility for biomedical and clinical risk stratification in population-based precision health research. These findings highlight the potential of simple and interpretable ML models to improve early PCa risk assessment in Hispanic men.