4P Machine learning-based prediction of survival in patients with metastatic renal cell carcinoma receiving first-line immunotherapy

Background
Immunotherapy has significantly improved survival outcomes for patients with metastatic renal cell carcinoma (mRCC). However, considerable variability exists in patient survival. Machine learning (ML) models offer an opportunity to harness diverse patient data to predict survival outcomes and tailor treatment strategies. This study aimed to develop ML models to predict the overall survival (OS) in mRCC patients receiving first-line immunotherapy.

Methods
We analysed 4895 mRCC patients from the National Cancer Database who received first-line immunotherapy since 2015. Fifteen features were selected based on the univariate Cox regression for OS, including demographics, Charlson-Deyo Score, tumour side, grade, lymph vascular invasion, and prior surgery or radiotherapy. Missing values were imputed using K-Nearest Neighbors. The data was split into training (70%) and testing (30%) sets. Classification and regression models were compared using hyperparameter tuning and 5-fold cross-validation. The SMOT technique addressed class imbalance.

Results
The 1-year and 3-year OS were 32.7% and 9.9%, respectively. Among the classification models, CatBoost demonstrated the best performance, with an area under curve (AUC) of 0.87, followed by LightGBM (0.86), XGBoost (0.86), and Decision Tree (0.86). The Decision Tree model achieved the highest F1 score (0.57), indicating a good balance between precision and recall. However, simpler models like Naive Bayes showed lower performance across all metrics. In the regression task, CatBoost also achieved the best performance, with a Mean Squared Error (MSE) of 115.5 and an R2 score of 0.52, indicating robust predictive accuracy. Feature importance analysis showed that tumour grade was the most significant predictor, followed by prior surgery and patient age. Socioeconomic factors, such as insurance status and facility type, also contributed significantly to the outcomes, while race had minimal predictive importance in this cohort.

Conclusions
Ensemble methods, particularly CatBoost, show superior performance in predicting mRCC outcomes. Tumour grade, surgery, and patient age emerged as key predictors.