Integrating graph neural networks with urine metabolomics for non-invasive detection of bladder cancer: a multi-centre case–control diagnostic study

Bladder cancer remains a major global health burden, and the stage at which a tumour is found largely determines what follows for the patient. Cystoscopy and urine cytology, the current mainstays, are invasive or insensitive, which keeps the search for accurate non-invasive alternatives open. We therefore asked a narrower question than a screening trial would ask: can a graph-based model separate treatment-naive bladder cancer patients from healthy controls using urinary metabolomic profiles alone? Liquid chromatography–mass spectrometry data were acquired for 338 participants (186 patients, 152 controls) at three centres, and 847 reproducible metabolite features passed quality control. Metabolite interaction graphs encoding partial correlations and KEGG pathway co-membership were estimated on training participants only, and a graph attention network (GAT) fusion architecture was trained on the resulting patient-level graphs. Every data-dependent step—quality-control filtering, normalisation, graph estimation, hyperparameter search and selection of the operating threshold—was confined to training and validation data, so that the internal test set and the external cohort were touched once. At the prespecified threshold the model reached an AUROC of 0.946 (95% CI 0.887–0.987), with sensitivity 0.919 (0.781–0.983) and specificity 0.900 (0.735–0.979) in 67 held-out participants, and an AUROC of 0.931 (0.873–0.977) in an independent cohort of 89 participants from a fourth centre. The model ranked above all conventional baselines, although with this sample size none of the pairwise differences survived correction for multiple comparisons, and a control model trained on randomised graph topology performed almost as well. Attention weights, permutation importance, integrated gradients and feature ablation converged on tryptophan catabolism, nucleotide biosynthesis and lipid remodelling. Because cases were histologically confirmed and controls were disease-free volunteers, these results describe case–control diagnostic discrimination; they do not establish detection before clinical presentation, and prospective enrolment of asymptomatic individuals remains necessary before any screening claim can be made.