Kidney cancer (KC) is diagnosed in approximately 350,000 people worldwide each year. While many patients with localized kidney cancer have favorable outcomes, roughly 20% already have distant metastasis at the time of diagnosis, and once cancer spreads, the 5-year survival rate drops dramatically to approximately 10%.
The lungs are the most common site where kidney cancer spreads, accounting for 55% of all metastatic cases. Patients who develop lung metastasis (LM) have a median survival time of only 15 months. Early identification of patients at high risk for lung metastasis could allow doctors to adopt preventive strategies, consider adjuvant therapies, or plan more aggressive treatment before metastasis occurs.
Existing tools like CT scans have relatively limited accuracy for predicting lung metastasis (62% sensitivity), and more advanced tests like MRI or biopsy are expensive and time-consuming. This study aimed to build a machine learning model using readily available clinical information that could predict lung metastasis risk with high accuracy before it occurs.
This large population-based study enrolled 52,714 patients with kidney cancer diagnosed between 2004 and 2017 from the SEER database (the United States Surveillance, Epidemiology, and End Results program, covering about 30% of the US population). Among these patients, 2,618 (about 5%) had developed lung metastasis at diagnosis.
The dataset was split into a training set (70% of patients) and an internal test set (30%). Critically, an external validation cohort of 492 patients from Southwest Hospital in Chongqing, China was also used to test the models on a completely independent population from a different country -- a key step in demonstrating that the model generalizes beyond the original dataset.
Clinical variables included in the analysis were: age, sex, race, T stage (how far the primary tumor has grown), N stage (whether nearby lymph nodes are involved), tumor size, histology (cancer cell type), and tumor grade (how abnormal the cells look). All of these are routinely available at diagnosis without additional testing.
Six machine learning algorithms were compared: Logistic Regression (LR), Extreme Gradient Boosting (XGB), Random Forest (RF), Support Vector Machine (SVM), Artificial Neural Network (ANN), and Decision Tree (DT). Each was optimized using 10-fold cross-validation.
The XGB algorithm significantly outperformed the others, achieving an AUC (Area Under the ROC Curve) of 0.913 in internal testing and 0.904 in external validation -- meaning it correctly distinguished patients with and without lung metastasis about 91% of the time. It also showed 87.3% sensitivity (correctly identifying patients who developed LM) and 80.9% specificity.
Decision curve analysis confirmed that XGB had the highest clinical applicability across a range of risk thresholds. The model's probability predictions clearly separated LM patients from non-LM patients, and Kaplan-Meier survival analysis confirmed that patients predicted by the model to have LM had significantly shorter survival times, validating the model's clinical relevance.
Feature importance analysis identified the top risk factors for lung metastasis across the three best-performing models (XGB, ANN, and LR). Consistently across all three models, the most important predictors were: T stage, N stage, tumor size, and tumor grade.
Patients with T stage 4 tumors (most locally advanced) had dramatically higher lung metastasis rates. Statistical analysis showed T4 tumors were associated with approximately 79 times higher odds of lung metastasis compared to T1 tumors in univariate analysis. N1 (lymph node positive) status increased odds by about 17 times.
Older age, male sex, larger tumor size, and higher tumor grade were independently associated with greater lung metastasis risk. Interestingly, Asian patients had higher lung metastasis rates than Black patients in this dataset, and tumor histology also played a role -- with some rare subtypes having substantially different metastasis patterns.
To make the XGB model practically usable, researchers built a free web-based online calculator available at: https://medicalmachinelearning.shinyapps.io/ModelForLungMetastasis/. Doctors and researchers can enter a patient's clinical information and instantly receive a prediction of lung metastasis risk.
Using this tool, for example, a female patient aged 60 with Asian ethnicity, T1-N0 stage, 44mm tumor size, transitional cell carcinoma histology, and grade I would be predicted as low risk for lung metastasis. This type of tool allows for individualized risk stratification at the time of diagnosis without any additional testing.
For high-risk patients identified by the model, clinicians might consider more intensive surveillance imaging, earlier referral to oncology for adjuvant therapy discussions, or participation in clinical trials targeting high-risk populations. Conversely, low-risk patients might be spared the cost and anxiety of aggressive surveillance protocols.
This study demonstrates that machine learning models built from routine clinical variables can predict lung metastasis in kidney cancer patients with high accuracy -- outperforming traditional clinical staging systems and prior models. The XGB model maintained strong performance even when validated on a geographically distinct patient population in China.
For patients newly diagnosed with kidney cancer, this type of risk model could provide important prognostic information at the very beginning of their care journey. Knowing whether you are at high or low risk for lung metastasis can help guide discussions with your care team about the urgency and aggressiveness of treatment.
The study's key limitations include the retrospective nature of the data, the lack of information about molecular features or treatment details in the SEER database, and the relatively small size of the external validation cohort (492 patients). Future studies should incorporate additional biomarkers and validate in larger, more diverse external populations. Despite these limitations, the model represents a meaningful advance in personalized kidney cancer risk prediction.