Standard detection methods have significant limitations. Cystoscopy, the current gold standard for bladder cancer detection, is invasive, expensive, and may miss tumors in certain areas of the bladder. Urine cytology, while noninvasive, has poor sensitivity especially for low-grade tumors.
Urine is in direct contact with bladder epithelial cells, making it a biologically logical medium for detecting cancer-related metabolic changes. Metabolites released from bladder cancer cells are expected to accumulate in urine at higher concentrations than in blood or other fluids.
Previous urine metabolomics studies using NMR, LC-MS, and GC-MS platforms have identified various putative markers including taurine, glycolysis-related compounds, and lipid metabolites -- but results have been inconsistent across laboratories due to differences in patient populations, instrumentation, and control group selection.
This study applied ultra-performance liquid chromatography time-of-flight mass spectrometry (UPLC-TOF-MS) to profile urine metabolomes from 87 bladder cancer patients and 65 hernia patients, combining statistical analysis with a decision tree machine learning model to identify discriminative metabolite markers.
152 patients were enrolled across two groups. The bladder cancer group comprised 87 patients (54 male, 33 female, average age 68.2 years), including 55 with early-stage non-muscle-invasive tumors and 32 with advanced muscle-invasive tumors. The control group consisted of 65 hernia patients (62 male, 3 female, average age 64.6 years).
Hernia patients were selected as controls because they underwent comparable procedures for urine sample collection -- first morning samples collected after admission and before surgical intervention -- minimizing collection-related confounders.
The 152 subjects were randomly divided into a training set of 105 samples (55 bladder cancer, 50 hernia) and an independent testing set of 47 samples (32 bladder cancer, 15 hernia). The training set was used for marker selection and model building, while the testing set served as a truly independent validation cohort.
Urine samples were processed by centrifugation, normalized by osmolality to 250 mOsm/kg, and prepared through methanol extraction before UPLC-TOF-MS analysis. Quality control samples were injected every 10 runs to ensure instrument stability throughout the batch analysis.
A four-step screening pipeline was applied to identify marker candidates. From 944,219 spectral ions detected per sample, the pipeline filtered ions by: (1) detection in more than half of training samples, (2) significant upregulation based on Gaussian-fitted fold-change distribution, (3) statistical significance by Wilcoxon rank-sum test (p-value less than 0.05), and (4) area under the ROC curve of at least 0.70.
This systematic filtering approach was designed to focus on upregulated markers specifically, as downregulation was considered potentially attributable to instrument detection limits rather than true biological absence. This conservative approach reduced the risk of false-positive marker identification.
The six surviving marker candidates were used to build a decision tree classifier using the C4.5 algorithm (implemented as J48 in the WEKA data mining toolkit). Decision trees are interpretable models that classify samples by asking a series of threshold-based questions about feature values.
5-fold cross-validation on the training set evaluated the stability and generalizability of the decision tree before final evaluation on the independent test set. The model was also validated visually using locally linear embedding (LLE) dimensionality reduction to confirm the discriminative power of the six markers in two-dimensional space.
Six spectral ions passed all four screening criteria from the pool of 944,219 detected ions. The six candidates had retention times ranging from 2.04 to 19.42 minutes and mass-to-charge ratios between 106.950 and 323.056. All six were significantly upregulated in bladder cancer samples compared to hernia controls.
Two of the six candidates -- eluting at 3.65 minutes with m/z values of 165.007 and 183.018 -- were identified as imidazoleacetic acid through database matching (Metlin and HMDB) and confirmation by product ion spectra. The remaining four candidates could not be definitively identified and represent novel unknown metabolites.
Locally linear embedding clustering of the six markers showed a clear separation between bladder cancer and hernia sample profiles, visually validating that the selected marker panel carries discriminative information even in reduced two-dimensional space.
The marker with the highest expression fold change was at 12.53 minutes with m/z 194.117, showing an 11.2-fold elevation in bladder cancer samples relative to hernia controls, though the biological identity of this compound remains uncharacterized.
5-fold cross-validation showed stable and consistent performance. Across five cross-validation iterations, the decision tree achieved average accuracy of 84.76% (standard deviation 1.75%), average sensitivity of 81.82% (standard deviation 1.61%), and average specificity of 88.00% (standard deviation 2.74%). Low standard deviations indicated minimal overfitting.
When evaluated on the independent test set of 47 previously unseen samples, the decision tree achieved an accuracy of 76.60%, sensitivity of 71.88%, and specificity of 86.67%. The higher specificity relative to sensitivity suggests the model is more reliable at correctly identifying non-cancer (hernia) samples than at catching all bladder cancer cases.
Performance dropped slightly from cross-validation to independent test evaluation, which is expected and reflects the real-world challenge of generalizing metabolomic models to new samples. The 76.60% accuracy on independent test data still represents a meaningful discriminative result given the small sample size.
The high specificity of 86.67% is clinically relevant as it suggests the marker panel would generate relatively few false positives -- an important consideration for any noninvasive screening test intended for use before more definitive invasive testing.
Imidazoleacetic acid has biological connections to bladder cancer. This metabolite is produced by the oxidation of histamine, which is primarily released by mast cells during inflammatory processes. Mast cells have been documented in association with bladder carcinoma, providing a plausible mechanistic link.
Upstream of imidazoleacetic acid, both histidine and histamine have been independently reported as potential bladder cancer markers. Histidine decarboxylase (HDC), which converts histidine to histamine, has been found expressed in other tumor types including melanoma and small cell lung carcinoma, suggesting a role in cancer-related metabolic reprogramming.
Prior metabolomic studies found elevated histidine in bladder tumor tissue compared to benign adjacent tissue, and further work has linked histidine to bladder cancer progression. The detection of imidazoleacetic acid in urine therefore appears consistent with upregulated histidine and histamine pathways driven by cancer-associated inflammation.
The divergence of marker signatures across studies reflects real methodological complexity. Differences in patient demographics, tumor stage distribution, sample collection timing, mass spectrometry platforms, and control group composition all contribute to variable metabolite signatures across laboratories, making standardization a key challenge for future biomarker validation.
Urine metabolomics shows promise as a noninvasive diagnostic tool for bladder cancer. The decision tree model built on just six metabolite markers achieved moderate accuracy and specificity on independent test data, supporting the concept that urine metabolite profiles can discriminate bladder cancer from non-cancer controls without invasive procedures.
The study had a relatively small sample size of 152 subjects, which limits the statistical power and generalizability of the findings. Validation in larger, independent cohorts with diverse patient populations will be essential before clinical application.
The four unidentified marker candidates represent an opportunity for further structural characterization. These novel metabolites may yield new biological insights into bladder cancer metabolism if their identities can be determined through advanced tandem mass spectrometry or isotope labeling experiments.
Integrating metabolomic markers with existing clinical tests such as cytology or protein biomarkers could potentially improve overall diagnostic accuracy. Future work should also explore whether the six-marker panel can distinguish bladder cancer stage or grade, extending its utility beyond initial detection to disease monitoring.