Renal cell carcinoma (RCC) is among the most common cancers globally, with approximately 431,000 new cases diagnosed in 2020 and nearly 180,000 deaths. Men are affected about 1.5 to 2 times more often than women, with incidence highest in older adults and in Western countries.
One of the most serious challenges with RCC is that it is almost always asymptomatic in its early stages. The classic triad of symptoms - blood in urine, flank pain, and a palpable abdominal mass - is now recognized as rare, occurring mainly in advanced disease. Most cases are found incidentally during imaging for other reasons.
By the time symptoms appear, the disease is often advanced - and advanced RCC has a very poor prognosis. This makes early detection critical. However, current screening programs for RCC are limited, and no validated, widely used blood or urine biomarker test exists for early kidney cancer detection.
The search for a simple, non-invasive test that can detect RCC early - from a urine or blood sample - has driven interest in liquid biopsy approaches. This study explores a laser-based technology called Surface-Enhanced Raman Scattering (SERS) applied to urine samples, combined with machine learning, as a potential screening tool.
Raman spectroscopy is a technique that shines a laser on a sample and measures how the light scatters. Different molecules scatter light in unique ways, creating a molecular 'fingerprint' that can identify what is in a sample. However, many molecules of interest in biological fluids are present in very tiny amounts, making their signals too weak to detect by standard Raman spectroscopy.
Surface-Enhanced Raman Scattering (SERS) solves this problem. When molecules are placed near the surface of tiny metallic nanoparticles - in this case, silver nanoparticles - the Raman signal is amplified by factors of up to one billion. This allows detection of trace amounts of molecules in complex biological fluids like urine.
A key advantage of SERS for cancer screening is that it is label-free - it does not require adding special dyes or markers to the sample. The urine is placed directly on a specialized solid substrate coated with silver nanoparticles, dried, and then analyzed with the laser. The entire process requires only 1 microliter of urine.
This study used a solid plasmonic substrate rather than colloidal nanoparticles suspended in liquid. This choice specifically enhances signals from small molecules (like urea, creatinine, and uric acid) rather than large proteins, making it ideal for capturing the metabolic fingerprint of urine that may differ between healthy individuals and cancer patients.
The study recruited 50 male patients with confirmed clear cell RCC (the most common kidney cancer subtype) and 44 healthy male donors. Urine samples were collected from cancer patients before any surgery or treatment, ensuring the SERS signal reflected the cancer state and not treatment effects.
Cancer patients ranged in age from 38 to 78 years. Of the 50 patients, 32 had Stage 1 cancer (localized, small tumors), 11 had Stage 2, and 7 had Stage 3 disease. This distribution of stages allowed the researchers to test whether SERS could also distinguish between different cancer stages - not just cancer versus no cancer.
Urine was stored at -80 degrees Celsius until analysis. For each sample, SERS spectra were systematically recorded at 50 different points across the dried urine droplet on the nanoparticle substrate. Two such maps were recorded per sample, and the 100 spectra were averaged to produce a single representative spectrum for each patient - improving reproducibility.
The key question was whether the SERS fingerprint of urine from cancer patients differed systematically from that of healthy individuals - and whether machine learning algorithms could learn these differences reliably enough to correctly classify new, unseen samples.
Statistical analysis revealed clear differences in the SERS spectra between cancer patients and healthy donors. Most notably, healthy donors showed significantly higher signals for urea (at 1004 cm-1) and creatinine (at 671 cm-1) - the two most abundant organic compounds in normal urine. Cancer patients had reduced levels of these metabolites.
This finding is biologically meaningful. Cancer disrupts normal metabolic pathways - including the urea cycle, the body's main way of eliminating nitrogen waste. Tumors redirect nitrogen metabolism toward building blocks for rapid cell growth, which reduces urea production and urinary urea levels. This is part of a recognized pattern called 'urea cycle dysregulation' seen in many cancers.
In contrast, cancer patients showed higher signals at certain other wavelengths (such as 390 cm-1 and 1335 cm-1), suggesting elevated levels of other metabolites. Some of these may correspond to ketone bodies or nucleic acid breakdown products - both of which can be altered by cancer metabolism.
The differences in individual peaks, while statistically significant, were not perfectly separating on their own. To correctly classify individual patients, a more sophisticated approach using all spectral information together was necessary - which is where multivariate analysis and machine learning came in.
Two main analytical approaches were applied to the SERS spectra. The first was Principal Component Analysis combined with Linear Discriminant Analysis (PCA-LDA) - a method that compresses the hundreds of data points per spectrum into a smaller set of components, then uses those components to draw a boundary between the cancer and healthy groups.
Using PCA-LDA with 13 principal components, the model achieved 100% accuracy in distinguishing RCC patients from healthy donors. Even with just six principal components specifically selected for their discriminating power, accuracy remained above 90%. These are remarkable figures that suggest the metabolic signal of kidney cancer in urine is strong and consistent.
The second approach, a Support Vector Machine (SVM), achieved 100% accuracy on the training set and 98.9 to 99% accuracy on validation data for distinguishing cancer from healthy urine. This near-perfect performance across two independent analytical methods greatly strengthens confidence in the findings.
When the researchers tested whether the model could also distinguish different cancer stages (I, II, and III), PCA-LDA achieved 88% overall accuracy - correctly classifying 83 of 94 samples. This is particularly promising because distinguishing early-stage from later-stage cancer could directly influence treatment decisions and follow-up intensity.
The near-perfect discrimination accuracy reported here is impressive but must be interpreted carefully. The study used leave-one-out cross-validation - a rigorous approach that tests each sample against a model built from all other samples. However, the study cohort of 94 participants is relatively small, and very high accuracy in small studies can sometimes reflect the specific characteristics of that group rather than a universal signal.
An important consideration is overfitting - when a model learns the quirks of a small dataset rather than genuine patterns. The researchers directly addressed this concern by showing that using fewer, more carefully selected principal components (those with the biggest differences between groups) achieved 90% accuracy with just 3 components - a simpler, more robust model that is less likely to overfit.
A practical limitation of current SERS technology for clinical use is that the equipment required - a high-resolution confocal Raman spectrometer - is expensive and found mainly in research laboratories. For this test to become a real clinical screening tool, point-of-care versions that are simpler and cheaper would need to be developed.
The study was also restricted to male patients with clear cell RCC. Whether the results apply equally to women, to other RCC subtypes, or to other kidney diseases is not yet known. Large, diverse validation studies - including comparisons with other kidney diseases that might also alter urine chemistry - are needed before clinical translation.
This study demonstrates that label-free SERS of urine combined with machine learning can discriminate kidney cancer patients from healthy donors with high accuracy. The approach is non-invasive, requires only a tiny urine sample, and does not require adding any chemical labels or markers to the sample.
The biological basis of the approach - differences in urea, creatinine, and other metabolites driven by cancer's disruption of normal metabolism - is well supported by the biochemical literature. This gives mechanistic credibility to the observed spectral differences beyond just statistical patterns.
Urine is an ideal diagnostic fluid for kidney cancer specifically: it is produced by the kidneys and may more directly reflect kidney tumor biology than blood. If this technology can be simplified and validated in larger populations, it could become part of a kidney cancer screening program that catches tumors at stages when they are still curable.
For patients and families, this research represents a vision of the future: a simple urine test, analyzed by AI, that could detect kidney cancer years before symptoms appear. Realizing that vision requires the next steps - larger studies, comparisons with other kidney diseases, testing in women and different ethnic groups, and engineering of practical low-cost devices.