Urine-Based FTIR Spectroscopy and Machine Learning Enable Non-Invasive Kidney Cancer Detection.

This study investigates the potential of Fourier transform infrared (FTIR) spectroscopy combined with multivariate and machine-learning analysis for kidney cancer detection using urine samples. FTIR analysis revealed distinct biochemical differences between urine from healthy controls and kidney cancer patients, with significant alterations observed in bands related to NH stretching, CH stretching, protein and urea-associated vibrations, phosphates, and carbohydrate-related regions. Principal component analysis showing group separation primarily along PC1 (33.9% variance), while uniform manifold approximation and projection provided enhanced nonlinear discrimination. Multiple models achieving high classification performance. With first-derivative preprocessing, both random forest and support vector machines reached 100% accuracy (sensitivity: 100%, 95% CI: 92.7%-100%; specificity: 100%, 95% CI: 91.2%-100%; AUC: 1.000 (95% CI: 1.000-1.000)). With raw spectral data, random forest achieved 98.9% accuracy (sensitivity: 100%, 95% CI: 92.7%-100%; specificity: 97.5%, 95% CI: 87.1%-99.6%; AUC: 1.000). Feature stability identified reproducible wavenumbers as cancer biomarkers.