Artificial Intelligence‐Based Urine Volatomic Assay for Multi‐Cancer Detection: A Multicentric Case–Control Discovery and Internal Validation Study
Cancer Reports · 2026-11 · 9 (11)
DOI: 10.1002/cnr2.70735 ↗Abstract
ABSTRACT Background Multi‐cancer detection assays may extend early‐detection efforts beyond organ‐specific screening programs and support emerging models of oncologic primary care, including age‐ and risk‐stratified multi‐cancer screening. We evaluated urine volatile organic compound profiles (volatome), an emerging biomarker source, for its association with cancer detection. Objective To determine whether urinary volatome profiles measured by gas chromatography–mass spectrometry (GC–MS) and interpreted with artificial intelligence (AI) distinguish participants with known cancer from cancer‐free controls and support exploratory tissue‐of‐origin (TOO) and stage classification. Methods This multicenter retrospective case–control discovery study included urine from 491 treatment‐naive participants with one of 10 cancers each and 200 cancer‐free healthy controls. Three participating sites contributed de‐identified specimens. Volatome profiles were measured by GC–MS. A proprietary ensemble tree‐based AI pipeline underwent repeated stratified fivefold internal cross‐validation. Preprocessing was fitted within each training partition. Each participant contributed one sample and was assigned only once within a given fold partition. The reported cancer‐signal operating point targeted 98% specificity. TOO analysis identified the most likely cancer type and its presumed tissue of origin. Results The cancer‐signal classifier produced an out‐of‐fold area under the receiver operating characteristic curve (AUC) of 0.990 (95% confidence interval [CI], 0.984–0.996). At the reported 98% specificity‐optimized cutoff, sensitivity was 91.9% (451/491). The positive predictive value (PPV) was 99.1% with the cohort's 71% (491/691) cancer prevalence. Grouped TOO accuracy was 91.6% (450/491; 95% CI, 89.2%–94.1%). Among 374 cancers with a recorded stage, the early‐versus‐late detection classifier had an AUC of 0.866, and it correctly assigned 98.1% (252/257) of early‐stage cancers to the early class while recovering 25.6% (30/117) of the late‐stage cancers. Recursive removal with high feature importance values produced a small, gradual decline rather than an abrupt loss of performance across the evaluated classification tasks. Conclusions Urinary VOC profiles contained a composite signal that discriminated known cancer cases from cancer‐free controls under internal cross‐validation with signals that were able to identify TOO with reasonable certainty. The feature‐removal analysis was consistent with distributed predictive information rather than reliance on one dominant VOC or a small feature panel. Prospective blinded validation in the intended‐use population remains necessary to establish population‐screening accuracy, comparative superiority, or clinical utility.
EUVIMED