Explainable multi-biomarker machine learning for prostate cancer risk stratification

Classification of prostate cancer risk is a clinical problem because the disease is heterogeneous and single-modal diagnostic approaches are limited. This paper constructed and tested an explainable multi-biomarker machine learning framework combining both clinical variables with imaging-based and molecular biomarkers to further enhance clinically meaningful prostate cancer. This was to focus on predictive performance, explainability, and reproducibility to facilitate clinical applicability. Several machine learning models were developed and tested with the help of standardized preprocessing and cross-validation techniques. The discrimination measures, calibration analysis, and decision-curve analysis were applied as a measure of clinical utility. The method of explainability was used to determine the influential biomarkers on the global and patient-specific scales, making the predictions easy to understand.
The findings revealed that multi-biomarker models were always better than unimodal methods in providing better discrimination and an increased net clinical benefit at pertinent decision thresholds. Imaging-derived and molecular characteristics had additional predictive value of clinical variables, which emphasized the significance of integrating multimodal predictive value. The explainability analyses showed biologically and clinically reasonable feature contributions and gave specific explanations that aided in interpreting individual-level risk predictions. Calibration of the model showed that the model agreement with the prediction and outcomes was quite good, and the accuracy of the model was consistent in all the folds of validation which supported the strength and reproducibility of the framework.