EXPLAINABLE MACHINE LEARNING FOR NON-INVASIVE PROSTATE CANCER DETECTION USING SIGNAL-DERIVED VOC FEATURES

Vollständiger Abstract

Worum geht es in dieser Arbeit?

The present study aimed to develop an explainable artificial intelligence (AI) model by using urinary volatile organic compound (VOC) profiles for non-invasive prostate cancer (PCa) detection. A total of 123 participants were included: 29 with PCa, 67 with benign prostatic hyperplasia (BPH), and 27 healthy controls. Urinary VOC signals were recorded by metal oxide semiconductor sensors across six thermal cycles, generating 120 features per subject. Mutual Information and Random Forest importance identified 15 key features. Five classifiers-logistic regression (LR), decision tree (DT), random forest (RF), naïve Bayes (NB), and ensemble voting (EV) were optimized with GridSearchCV and evaluated by using stratified cross-validation with SMOTE. The NB classifier provided the best discriminative performance on the independent test set, achieving an ROC-AUC of 0.90, while its accuracy, precision, recall, and F1-score were comparable to those of LR and RF under stratified 5-fold cross-validation. Importantly, explainability in this model arises from the probability-based decision structure of NB, which generates two distinctly separated probability clusters for PCa and non-PCa. This bimodal distribution enables direct interpretation of risk estimates, offering transparent and easy-to-communicate clinical meaning an advantage over black-box models. LR showed comparable utility (ROC-AUC = 0.81), whereas RF (0.76) and EV (0.77) demonstrated moderate discriminative performance, and DT produced the weakest results (0.67). These findings indicate that urinary VOC-derived features, analyzed with interpretable AI, can differentiate PCa from benign and control cases with reproducible performance across cross-validation folds. NB emerged as the most robust and transparent approach in this study, supporting the potential of VOC-based AI as a low-cost adjunct to clinical assessment and diagnostic pathways. Multi-center validation is recommended to confirm generalizability.