Improved predictive performance of prostate biopsy collaborative group risk calculator when based on automated...

Area of Science:

• Urology
• Oncology
• Machine Learning in Healthcare

Background:

• The Prostate Biopsy Collaborative Group risk calculator (PBCG RC) demonstrates moderate accuracy in predicting prostate cancer (PCa).
• Existing risk calculators require enhancement for improved diagnostic capabilities.

Purpose of the Study:

• To develop an automated machine learning (AutoML) enhanced PBCG RC for predicting any-grade and high-grade prostate cancer (PCa).
• To improve the discriminatory capability of the original PBCG RC using AutoML.

Main Methods:

• A retrospective single-center study involving 832 patients undergoing prostate biopsy with PSA levels between 2-50 ng/ml.
• Utilized H2O, an open-source AutoML platform, training 20 base learning algorithms.
• Compared the developed AutoML PBCG RC against the original PBCG RC for discrimination, calibration, and clinical utility.

Main Results:

• Prostate cancer (PCa) was detected in 41% of patients, with 19.1% having high-grade PCa.
• AutoML models showed superior discrimination for PCa (AUC: 0.703 vs 0.628) and high-grade PCa (AUC: 0.990 vs 0.717) compared to the original PBCG RC.
• Decision curve analyses indicated better performance of AutoML models, with PSA identified as the most crucial feature for high-grade PCa prediction.

Conclusions:

• Ensemble techniques and AutoML algorithms were employed to create a publicly accessible online PCa risk tool.
• The AutoML models substantially enhanced the predictive performance of the original PBCG RC, achieving near-perfect predictions for high-grade PCa.
• External validation is recommended before widespread clinical adoption of the new models.