Prostate cancer treatment recommendation study based on machine learning and SHAP interpreter

Area of Science:

• Oncology
• Medical Informatics
• Biostatistics

Background:

• Prostate cancer treatment decisions require accurate prediction models.
• Existing models may lack transparency or comprehensive validation.

Purpose of the Study:

• To develop and validate machine learning models for predicting prostate cancer treatment options (surgical vs. non-surgical).
• To identify key factors influencing treatment prediction.
• To compare survival rates between different treatment modalities.

Main Methods:

• Utilized data from 140,294 prostate cancer cases from the Surveillance, Epidemiology, and End Results (SEER) database.
• Applied 10 machine learning algorithms, evaluating performance using AUC, accuracy, sensitivity, and specificity.
• Employed Shapley Additive Explanations (SHAP) for factor importance and survival analysis for outcome comparison.

Main Results:

• The CatBoost model achieved the highest performance (AUC=0.939, accuracy=0.877).
• Key predictors included T stage, cancer stage, age, cores positive percentage, prostate-specific antigen (PSA), and Gleason score.
• Surgery improved 10-year survival by 20.36% compared to non-surgical treatments; radical prostatectomy showed the highest 10-year survival (89.2%).

Conclusions:

• Developed a robust, transparent predictive model to aid prostate cancer treatment decisions.
• Identified critical factors for personalized treatment planning.
• Demonstrated the survival benefit of surgical intervention, particularly radical prostatectomy.