Development of Machine Learning Models for Predicting Prostate Cancer in Biopsy Candidates Using Prostate-Specific...

Area of Science:

• Oncology
• Radiology
• Machine Learning

Background:

• Prostate-specific antigen (PSA) alone is insufficient for prostate cancer (PCa) diagnosis, especially in the 4-10 ng/mL range.
• Multiparametric magnetic resonance imaging (mpMRI) has limitations in the PCa diagnostic pathway.
• Hematologic parameters offer additional insights for PCa detection.

Purpose of the Study:

• Develop a biopsy prediction model for prostate cancer.
• Evaluate machine learning (ML) algorithms incorporating PSA, mpMRI, and hematologic data.
• Improve diagnostic accuracy for PCa in patients with PSA levels ≤10 ng/mL.

Main Methods:

• Included patients with PSA ≤10 ng/mL who underwent biopsy (2017-2024).
• Applied five ML techniques: logistic regression, random forest (RF), extra trees, XGBoost, and light GBM.
• Utilized PSA levels, mpMRI findings, and hematologic parameters for model development.

Main Results:

• XGBoost model achieved the highest accuracy with 94.74% sensitivity and 100% specificity.
• Random Forest (RF) model showed 78.95% sensitivity and 100% specificity.
• Key predictors included free/total PSA, PI-RADS score 4, and platelet-to-lymphocyte ratio.

Conclusions:

• XGBoost and RF models demonstrate excellent performance for PCa biopsy prediction.
• High AUC values (0.97 for XGBoost, 0.89 for RF) indicate strong predictive power.
• External validation is recommended to generalize ML model accuracy across diverse populations.