Robust Logistic Regression-based Diagnosis Method of Prostate Cancer Using Optimized Feature Selection on Race...

Prostate cancer (PCa) incidence varies significantly by race, with Black men experiencing nearly 1.8 times higher prevalence than White men in the USA. Current prostate specific antigen (PSA)-based diagnostics lack specificity, and many machine learning models fail to consider racial differences in gene expression. This study proposes a race-aware PCa detection framework using optimized feature selection to improve diagnostic accuracy and fairness.

RNAseq-Count-STAR and clinical phenotype data from TCGA (554 patients) were analyzed. A feature selection pipeline integrating Differential Gene Expression analysis, Receiving Operating Characteristic (ROC) analysis, and Gene-Set Enrichment Analysis identified a 9-gene subset strongly associated with the PCa clinical pathway. The model was trained on White population data and validated on the Black population dataset using various data balancing techniques.

The 9-gene logistic regression model achieved 95% accuracy in the White population and 96.8% accuracy in the Black population. Fairness analysis indicated minimal disparity between groups (4% difference in demographic parity, p=0.518). These results highlight the predictive value of race-specific biomarkers and demonstrate that biologically informed feature selection improves both accuracy and interpretability.

This study introduces a race-specific PCa detection framework that improves diagnostic accuracy using targeted biomarkers. It addresses misclassification risks in race-agnostic models and emphasizes the need for race-aware gene expression in ML diagnostics. Beyond detection, it enables personalized treatment, advancing precision medicine in PCa care.