Urinary volatilomics using liquid-liquid extraction and gas chromatography-mass spectrometry (GC-MS) combined with machine learning algorithms as a tool for diagnosis and surveillance of urothelial bladder cancer.

Abstract

Background: Bladder cancer is the 11th most common cancer in the United Kingdom, with approximately 10,500 new cases annually. Diagnosis and surveillance typically involve cystoscopy, an expensive, time-consuming, and uncomfortable procedure which has encouraged efforts to identify biomarkers, particularly in urine, given its direct contact with malignant tissue.

Methods: Urine collected from 100 participants (50 bladder cancer patients, 50 controls) was subjected to solvent extraction followed by gas chromatography-mass spectrometry (GC-MS) to determine potential volatile and semi-volatile biomarkers. The results were analysed using classical univariate statistics and machine learning methods. Five machine learning algorithms were evaluated, with recursive feature elimination (RFE) identifying optimal biomarker panels.

Results: Machine learning with XGBoost achieved area under the receiver operating characteristic curve (AUROC) of 0.869 (95% CI: 0.740-0.988), representing a significant improvement over the classical statistical approach (AUROC 0.752). An 8-metabolite panel achieved balanced sensitivity and specificity of 85%, or 95% sensitivity with 70% specificity when optimised for screening.

Conclusions: The findings indicate that solvent extraction of urine shows promise for isolating putative biomarkers of bladder cancer. Employing machine learning achieved diagnostic accuracy potentially suitable for clinical deployment as a non-invasive bladder cancer detection tool.