A gradient boosting machine-based approach to predicting chemotherapy response in patients with metastatic colorectal cancer

Identifying predictors of response or progression after first-line chemotherapy in metastatic colorectal cancer is challenging. In this study, we attempted to predict patient outcomes using various machine learning methods. We recruited all patients diagnosed with metastatic colorectal cancer (mCRC) at an academic center within a specified time period. Response to first-line chemotherapy and associated factors were assessed using various machine learning models. A total of 101 newly diagnosed mCRC patients starting first-line chemotherapy were included, with a median age of 62 and 69% male. We tested 15 machine learning models for predicting the best response to chemotherapy, with Light GBM showing the highest initial accuracy of 0.71. After tuning, LightGBM’s accuracy increased to 0.79, with an AUC of 0.82. The most important features were age at diagnosis, maximum metastatic dimension, and metastatic status at diagnosis. Genetic variables did not significantly impact the response prediction. However, model performance decreased on the test dataset (accuracy: 0.60), indicating potential overfitting. Using LightGBM, we developed a predictive model demonstrating moderate performance in estimating response to first-line chemotherapy in mCRC patients. However, model performance decreased on the test dataset (accuracy: 0.60), indicating potential overfitting. While the model showed potential in identifying factors associated with treatment response, its clinical applicability remains limited and requires validation in larger, independent cohorts.