Integrating radiological and clinical data for clinically significant prostate cancer detection with machine learning...

Area of Science:

• Urology
• Oncology
• Medical Imaging
• Machine Learning

Background:

• Prostate cancer (PCa) risk prediction often relies on clinical data and magnetic resonance imaging (MRI) using the Prostate Imaging-Reporting and Data System (PI-RADS).
• Current models predominantly use logistic regression, prompting exploration of advanced machine learning techniques for improved accuracy.

Purpose of the Study:

• To compare the efficacy of various machine learning models against logistic regression in predicting clinically significant prostate cancer (CsPCa).
• To evaluate the clinical utility of these models in reducing unnecessary biopsies.

Main Methods:

• A dataset of 4799 patients from Catalonia, Spain, was used with an 80-20% train-validation split.
• Predictor variables included age, PSA, prostate volume, PSAD, DRE, family history, prior negative biopsy, and PI-RADS categories.
• Models evaluated: logistic regression, ridge, LASSO, elastic net, classification trees, random forest, XGBoost, and neural networks.

Main Results:

• XGBoost demonstrated the highest specificity (0.640) at 0.9 sensitivity, closely followed by random forest and neural networks.
• XGBoost also showed the greatest potential to avoid unnecessary biopsies (41.77%), outperforming logistic regression (40.62%).
• SHAP analysis identified PI-RADS (especially 4 and 5) as the most influential factor, with DRE and family history also significant, and prior negative biopsy being protective.

Conclusions:

• Advanced machine learning models, particularly XGBoost, offer superior predictive performance and clinical utility for CsPCa detection over traditional logistic regression.
• These models can enhance decision-making, potentially reducing the number of invasive and unnecessary prostate biopsies.
• PI-RADS scores, combined with clinical factors, are crucial for accurate risk stratification, with machine learning providing a more nuanced approach.