Clear cell renal cell carcinoma (ccRCC) is the most common histologic subtype of kidney cancer and is associated with poor prognosis when diagnosed at advanced stages. Most patients have no symptoms until tumors are large and have spread beyond the kidney, making early detection challenging under current practice, which relies primarily on incidental imaging findings.
Blood-based biomarkers offer an attractive approach for population-level screening because blood draws are minimally invasive, widely accessible, and routinely performed as part of standard clinical care. If specific combinations of routine blood test values could reliably discriminate ccRCC from healthy controls, they could be incorporated into existing laboratory panels without requiring new or expensive assays.
This study evaluated whether an 8-indicator panel of routine blood biochemistry tests, combined with a random forest machine learning model, could accurately screen for early ccRCC. The panel was designed to be practical for widespread clinical implementation, using only biomarkers available through standard complete blood count and metabolic panel testing.
The study included 743 ccRCC patients and 500 healthy controls, providing a sufficiently powered cohort for model training and evaluation. The ccRCC patients were confirmed by histopathology following surgical resection, ensuring that the positive cases represent true cancer rather than radiologically suspected masses. Healthy controls were drawn from individuals without cancer history and with normal imaging.
The eight biomarkers selected for the panel were albumin (ALB), total protein (TP), hemoglobin (HB), potassium (K+), blood urea nitrogen (BUN), creatinine (CREA), uric acid (UA), and high-sensitivity C-reactive protein (hs-CRP). These markers were chosen because they reflect metabolic, inflammatory, and nutritional processes that are known to be altered in malignancy.
Two derived biomarker ratios were also analyzed: the albumin-to-globulin ratio (AGR) and the BUN-to-creatinine ratio (BCR). These ratios can capture proportional changes in metabolic markers that may be more diagnostically sensitive than any single marker in isolation. The AGR in particular has been identified as a cancer-associated inflammatory biomarker in multiple tumor types.
A random forest classifier was trained on the eight-biomarker panel to distinguish ccRCC from healthy controls. Random forests are ensemble learning methods that build many decision trees on random subsets of the training data and aggregate their predictions, producing a classifier that is more robust and generalizable than any single tree.
Model performance was assessed using area under the receiver operating characteristic curve (AUC), sensitivity, and specificity. The dataset was split into training and test sets, with care taken to evaluate performance on held-out data to provide an unbiased estimate of generalization performance.
In addition to the combined random forest model, the predictive value of each individual biomarker was evaluated separately, allowing identification of which single markers carry the most discriminatory information and which add incremental value only in combination with others.
The random forest model combining all eight biomarkers achieved an AUC of 0.932, indicating excellent discriminative ability between ccRCC patients and healthy controls. Sensitivity was 88.2% and specificity was 86.3%, representing a clinically meaningful balance between correctly identifying cancer patients and correctly clearing healthy individuals.
At 88.2% sensitivity, the model would detect approximately 882 out of every 1,000 true ccRCC cases in a screened population. The 86.3% specificity means that approximately 137 out of every 1,000 healthy individuals would receive a false positive result, requiring follow-up imaging. These performance characteristics compare favorably with many established cancer screening tests.
The multi-biomarker combination substantially outperformed any single marker, confirming that the eight biomarkers provide complementary and partially independent discriminatory information. This is consistent with the known multidimensional biology of ccRCC, which alters metabolism, inflammatory signaling, kidney function, and nutritional status simultaneously.
Among all individual biomarkers, hs-CRP (high-sensitivity C-reactive protein), a systemic marker of inflammation, achieved the highest single-marker AUC of 0.873. This finding is consistent with the well-established role of cancer-associated systemic inflammation in RCC biology, where elevated CRP is associated with tumor-driven cytokine release and adverse immune modulation.
The AGR (albumin-to-globulin ratio) was also a strong individual predictor, reflecting the nutritional depletion and inflammatory protein shifts characteristic of malignancy. Low albumin combined with elevated globulins (which include acute-phase proteins) produces a low AGR that signals cancer-associated inflammation and nutritional compromise.
The BUN-to-creatinine ratio (BCR) reflects kidney functional reserve and protein catabolism. Alterations in BCR may reflect subclinical impairment of kidney function caused by the growing RCC tumor, providing a functional marker that complements the inflammatory and nutritional information captured by hs-CRP and AGR.
The practical appeal of this approach lies in its reliance on biomarkers that are already measured in millions of routine blood draws annually. Unlike specialized cancer biomarkers requiring dedicated assays, the eight markers in this panel are part of standard complete blood count and metabolic panels ordered for a wide range of clinical indications.
A validated random forest model applied to these routine results could flag patients warranting renal imaging without requiring any additional blood testing or novel biomarker assays. This makes the approach highly scalable and potentially cost-effective as a first-line screening trigger, particularly in populations at elevated risk such as smokers, obese individuals, and those with hypertension or familial RCC history.
Limitations include the case-control design, which does not reflect the low prevalence of ccRCC in general populations and likely inflates apparent sensitivity and specificity. Performance in true screening populations would need to be assessed prospectively. External validation in independent cohorts from different geographic regions and healthcare systems is also needed before clinical adoption. Despite these caveats, the study provides strong proof-of-concept that routine blood biochemistry can inform ccRCC risk stratification.