RCC-Supporter is a machine learning-based software tool designed to help doctors decide which treatment is most appropriate for a patient with renal cell carcinoma (RCC), the most common type of kidney cancer.
Treatment decisions in kidney cancer are complex because multiple options exist, including active surveillance (watchful waiting), surgical removal of the tumor (nephrectomy), and systemic chemotherapy. The best choice depends on dozens of patient and tumor characteristics that can be difficult to weigh simultaneously.
By analyzing real-world patient data, RCC-Supporter learns which combinations of factors are associated with which treatment outcomes, then uses that knowledge to make treatment recommendations for new patients. This type of tool is called a clinical decision support system (CDSS).
The system was built using records from 1,867 RCC patients treated at Pusan National University Yangsan Hospital in South Korea between 2008 and 2021. This large, real-world dataset gave the model diverse examples to learn from.
Patients were divided into three treatment groups based on what they actually received: surveillance (40%), surgery (48%), and chemotherapy (12%). The model was trained to predict which treatment category a new patient belongs to based on their clinical characteristics.
Multiple machine learning algorithms were tested, including Gradient Boosting Machines (GBM), Random Forest, Decision Tree, XGBoost, LightGBM, Logistic Regression, and K-Nearest Neighbors (KNN), to find the one that made the most accurate predictions.
The Gradient Boosting Machine (GBM) algorithm with 100 to 150 features achieved the highest accuracy, correctly predicting the appropriate treatment in 95% of cases, with a 95% confidence interval of 92% to 98%.
The model also achieved an area under the curve (AUC) of 0.87 in multi-class classification, meaning it reliably distinguished between the three treatment categories. This level of performance is strong enough to be considered clinically relevant.
The most influential features for making predictions were: whether RCC was diagnosed by CT scan, TNM staging (T and N stage), alkaline phosphatase, hemoglobin, red blood cell (RBC) count, and lactate dehydrogenase (LDH). These are all factors already routinely measured in kidney cancer patients.
Kidney cancer treatment decisions involve many variables that interact in complex, non-linear ways. A patient's age, kidney function, tumor size, cancer stage, blood test results, and other factors all play a role. Even experienced oncologists can find these decisions challenging.
A clinical decision support system does not replace the physician but acts as a second opinion based on large-scale pattern recognition across thousands of similar cases. This can be especially valuable in smaller hospitals or resource-limited settings where specialist expertise may be less available.
Machine learning tools can also identify patterns that are not intuitively obvious. For example, the finding that routine blood markers like alkaline phosphatase and LDH strongly influenced treatment classification highlights laboratory values that might be underweighted in clinical practice.
TNM staging refers to Tumor size, Node involvement, and Metastasis status, and is the standard system for classifying cancer severity. Higher T and N stages generally correlate with more aggressive disease and favor active treatment over surveillance.
Alkaline phosphatase (ALP) is a liver and bone enzyme that can be elevated when cancer spreads to these organs. LDH (lactate dehydrogenase) is a marker of cell turnover and tissue damage, often elevated in more advanced cancers.
Hemoglobin and RBC count reflect whether a patient has anemia, which is common in kidney cancer and can indicate disease burden or nutritional status. Low hemoglobin may signal that more aggressive treatment is needed sooner.
For patients, a tool like RCC-Supporter could provide reassurance that the recommended treatment aligns with patterns seen across thousands of similar cases. It supports shared decision-making by giving both doctors and patients a data-driven basis for treatment discussions.
The tool's reliance on standard clinical variables means it does not require expensive genomic testing to work. This makes it accessible in a wide range of healthcare settings, including those in lower-resource environments.
Future versions of the system could incorporate additional data types, such as imaging features or genomic markers, to further improve accuracy. The research team envisions this as an evolving tool that improves as more patient data becomes available.