Near-infrared spectroscopy for urine-based cancer detection: a systematic characterization of the analytical challenges of a complex biofluid matrix.

Near-infrared spectroscopy (NIRS) of urine offers an attractive route to non-invasive cancer screening. However, the analytical complexity of this matrix has not been systematically characterized. This study comprehensively evaluated the ability of NIRS to distinguish between the control group and patients with kidney or prostate cancer using urine samples (n = 85). Results revealed that 93.51% of total spectral variance reflects physiological and measurement sources unrelated to cancer, rendering conventional chemometric approaches, aquaphotomics, and individual wavelength feature selection ineffective. Cancer-discriminative information exists but is weak, buried in principal components that together explain only 0.004% (kidney cancer) to 0.39% (prostate cancer) of total spectral variance. Targeted feature selection on principal component analysis (PCA) scores identified these components and enabled machine learning classification. This approach achieved up to 78.8% balanced accuracy for kidney cancer and 75.1% for prostate cancer. Kidney and prostate cancer were associated with entirely non-overlapping principal components, suggesting distinct reorganizations of the urine water matrix for each cancer type. This finding reveals the fundamental physico-chemical constraints of urine for NIRS-based cancer detection: the water hydrogen bonding network produces spectral effects that dwarf the cancer-related perturbation. Thus, this study provides not a demonstration of clinical utility, but a systematic quantification of the analytical boundaries of urine NIRS for cancer detection.