Using Supervised Machine Learning Algorithms to Predict Prostate Cancer Cases: A Case of Men of African Descent Carcinoma of the Prostate Consortium (MADCaP), South Africa

Abstract

Prostate cancer is a major public health concern globally, particu- larly affecting men of African descent who experience higher incidence and mortality rates. This study aimed to develop predictive mod- els for prostate cancer risk using machine learning algorithms applied to data from the Men of African Descent Carcinoma of the Prostate (MADCaP) consortium at Chris Hani Baragwanath Academic Hospital (CHBAH). The study population included 1096 prostate cancer cases enrolled between January 2017 and December 2021. A retrospective cross- sectional analysis of secondary data was conducted. Using Python in Google Colaboratory, we implemented Na¨ıve Bayes, K-Nearest Neigh- bors (KNN), Decision Tree, Random Forest, and Support Vector Ma- chine (SVM) models. Data preprocessing followed the Knowledge Dis- covery in Databases (KDD) methodology, including cleaning, feature selection, and addressing class imbalance with a hybrid oversampling technique. Model performance was evaluated using accuracy, precision, recall, and F1 score. Results showed that KNN and Decision Tree models achieved the best predictive performance, with KNN providing high accuracy and Deci- sion Tree offering interpretability. Random Forest performed well but iii indicated minor overfitting. The study population analysis revealed significant differences in age across risk categories (H = 2631.90, p < 0.001) and a statistically significant change in PSA levels from referral to diagnosis (t = 3.20, p = 0.001). Other factors, such as smoking status, showed no significant association with risk classification. In conclusion, machine learning models, particularly KNN, show promise for predicting prostate cancer risk in men of African descent. These findings support targeted screening and risk stratification in high-risk populations. Future studies should validate these models across diverse datasets and consider integrating additional genetic and environmental factors to improve predictive accuracy and address health disparities.