Background: Prostate cancer (PCa) screening using serum prostate-specific antigen (PSA) testing has caused unnecessary biopsies and overdiagnosis owing to its low accuracy and reliability. Therefore, there is an increased interest in identifying better PCa biomarkers. Studies showed that trained dogs can discriminate patients with PCa from unaffected men by sniffing urine. We hypothesized that urinary volatile organic compounds (VOCs) may be the source of that odor and could be used to develop urinary VOC PCa diagnosis models.
Patients and methods: Urine samples from 55 and 53 biopsy proven PCa-positive and -negative patients respectively were initially obtained for diagnostic model development. Urinary metabolites were analyzed by gas chromatography-mass spectrometry. A PCa diagnosis model was developed and validated using innovative statistical machine-learning techniques. A second set of samples (53 PCa-positive and 22 PCa-negative patients) were used to evaluate the previously developed PCa diagnosis model.
Results: The analysis resulted in 254 and 282 VOCs for their significant association (P < .05) with either PCa-positive or -negative samples respectively. Regularized logistic regression analysis and the Firth method were then applied to predict PCa prevalence, resulting in a final model that contains 11 VOCs. Under cross-validation, the area under the receiver operating characteristic curve (AUC) for the final model was 0.92 (sensitivity, 0.96; specificity, 0.80). Further evaluation of the developed model using a testing cohort yielded an AUC of 0.86. As a comparison, the PSA-based diagnosis model only rendered an AUC of 0.54.
Conclusion: The study describes the development of a urinary VOC-based model for PCa detection.
Keywords: Gas chromatography-mass spectrometry; Metabolomics; Nonparametric statistics; Prostatic neoplasms; Stir bar sorptive extraction.
Copyright © 2019 Elsevier Inc. All rights reserved.