A machine learning approach to predict progression on active surveillance for prostate cancer

Area of Science:

• Urology
• Oncology
• Medical Informatics

Background:

• Active surveillance (AS) for prostate cancer requires robust prediction models to guide risk-adapted management.
• Current models for predicting AS progression primarily use traditional statistical methods.
• Machine learning (ML) offers a potential avenue to improve predictive accuracy.

Purpose of the Study:

• To evaluate if ML approaches can enhance the prediction of grade-progression in prostate cancer patients managed with AS.
• To compare the performance of various ML classifiers against traditional logistic regression for this prediction task.

Main Methods:

• A retrospective cohort study included 790 patients with very-low or low-risk prostate cancer on AS (1997-2016).
• Traditional logistic regression (T-LR) and several ML classifiers (SVM, Random Forest, ANN, ML-LR) were trained to predict grade-progression.
• Model performance was assessed using the F1 score in a test set.

Main Results:

• Out of 790 patients, 234 experienced grade-progression during a median follow-up of 6.29 years.
• ML models demonstrated superior performance, with the Support Vector Machine achieving the highest F1 score (0.586).
• All ML models significantly outperformed T-LR (F1 score 0.182) (p < 0.001).

Conclusions:

• ML methods significantly outperformed traditional logistic regression in predicting prostate cancer progression on AS.
• These findings suggest ML can facilitate more robust and individualized risk-stratification for patients on active surveillance.
• Further validation of these ML models is warranted for clinical implementation.