Leveraging readily available clinical data with machine learning to predict first-line immunotherapy outcomes in non-small cell lung cancer

Abstract

Background

Immune checkpoint inhibitors (ICIs) are essential first-line treatments for recurrent or metastatic non-small cell lung cancer (NSCLC). However, predicting their effectiveness and the occurrence of immunotherapy-related adverse events (irAEs) remains challenging.

Methods

This retrospective study involved NSCLC patients who received first-line ICI therapy at China-Japan Friendship Hospital in Beijing, China, between October 29, 2018, and July 10, 2024. We employed five machine learning models to predict treatment responses to ICIs and the occurrence of irAEs.

Results

A total of 397 NSCLC patients who received first-line ICIs were included in the analysis, with 277 patients in the train-validation cohort and 120 in the test cohort. The neural network and gradient boosting models were the most effective for predicting treatment responses, achieving AUC values of 0.87 and 0.84, respectively. For predicting irAEs, random forest and gradient boosting emerged as the top performers, with AUC values of 0.84 and 0.80. Feature importance analysis identified key predictors such as red blood cell (RBC) counts and metastatic sites for treatment response, while metastatic sites and sex were significant for irAE prediction. In the validation cohort, the neural network demonstrated strong performance in predicting treatment response (AUC of 0.84, recall of 0.8406, and F1 score of 0.8007), while the random forest model excelled in predicting irAEs (AUC of 0.82, accuracy of 0.7417, precision of 0.7500, recall of 0.8261, and F1 score of 0.7862).

Conclusion

These findings highlight the potential for enhancing personalized treatment strategies for NSCLC patients undergoing first-line ICI therapy.