Developing a Predictive Model for Significant Prostate Cancer Detection in Prostatic Biopsies from Seven Clinical...

Area of Science:

• Medical Imaging and Diagnostics
• Machine Learning in Healthcare
• Prostate Cancer Research

Background:

• Prostate cancer (PCa) diagnosis relies on accurate risk stratification.
• Existing predictive models like the Barcelona (BCN-MRI) model inform clinical decisions.
• Comparing advanced machine learning (ML) with traditional logistic regression (LR) is crucial for improving PCa detection.

Purpose of the Study:

• To compare the performance of ML and LR algorithms in predicting PCa.
• To evaluate a novel feedforward neural network (FNN)-based SimpleNet model (GMV) against the established BCN-MRI logistic regression (LR) model.
• To assess predictive accuracy, discrimination, precision-recall, and clinical utility for PCa detection.

Main Methods:

• Utilized a cohort of 5005 men suspected of PCa undergoing MRI.
• Developed and validated a SimpleNet (GMV) ML model and a logistic regression (BCN) model.
• Evaluated models using area under the curve (AUC), precision-recall metrics, and clinical utility assessments.

Main Results:

• Both GMV (ML) and BCN (LR) models demonstrated strong predictive performance (AUCs 0.88/0.85 for GMV, 0.85/0.84 for BCN).
• GMV model showed superior recall (sensitivity), while BCN model offered higher precision and specificity.
• Both models significantly reduced unnecessary prostate biopsies by approximately 27-29% while maintaining 95% sensitivity.

Conclusions:

• Machine learning and logistic regression models provide high accuracy for PCa detection.
• ML models offer enhanced sensitivity (recall), beneficial for ruling out disease.
• LR models provide higher specificity, useful for reducing unnecessary invasive procedures; model choice depends on clinical priorities.