Optimized Prostate Cancer Stage Classification Using XGBoost Based on Racial miRNA Expression

Prostate cancer has a high mortality rate and shows diagnostic disparities between racial groups, particularly between White and Black populations. This study aimed to develop a prostate cancer stage classification model using the XGBoost algorithm on miRNA expression data with a focus on race-based analysis.

The data used in this study were obtained from the Genomic Data Commons. The Cancer Genome Atlas through the UCSC Xena Browser, consisting of miRNA expression and patient clinical data. Several feature selection methods were applied, including Student's t-test, mutual information, chi-square, and limma. Data balancing techniques such as RandomOversampler, SMOTE, SMOTEENN, and BorderlineSMOTE were also implemented to address class imbalance.

The results showed that the XGBoost model achieved an accuracy of up to 89% on data from White patients. Additionally, the F1-score shows the results of 92%, these results suggest this model is particularly strong in identifying minority class in this unbalance data. However, when tested on data from Black patients, the accuracy decreased to 72-74%, indicating limitations in cross-race performance. Additionally, Local Interpretable Model-agnostic Explanations (LIME) was implemented to identify the gene features that contributed most significantly to the model's predictions.

The study demonstrates that XGBoost, combined with race-specific feature selection and hyperparameter tuning, can accurately detect prostate cancer with high performance. These findings highlight the potential of XGBoost for improving prostate cancer detection while also emphasizing the importance of considering race-specific differences in machine learning models.