Development and validation of a machine learning model for predicting objective response to PD-1 inhibitors in lung
Weiqi Chen1, Xiaojie Zhao2, Zhuojia Li1
1Department of Pharmacy, The Second Affiliated Hospital, Guangzhou Medical University, Guangzhou, China.
Background:
Given the substantial heterogeneity in the efficacy of immunotherapy for lung cancer, identifying reliable predictive factors is crucial. This study aims to develop a machine learning (ML) model to identify predictors of achieving objective response in lung cancer patients receiving programmed cell death protein 1 (PD-1) inhibitor therapy.
Methods:
A retrospective analysis was conducted on data from lung cancer patients treated with PD-1 inhibitors at The Second Affiliated Hospital of Guangzhou Medical University between November 2021 and November 2024. Patients were categorised into response group [complete response (CR) + partial response (PR)] and non-response group [stable disease (SD) + progressive disease (PD)] based on treatment efficacy. Following least absolute shrinkage and selection operator (Lasso) analysis to screen clinical characteristics, 10 ML algorithms were employed to evaluate the top 10 response-associated factors for model construction. Model performance was rigorously assessed using metrics including area under the receiver operating characteristic (ROC) curve (AUC), accuracy, sensitivity, specificity, and F1 score. The SHapley Additive exPlanations (SHAP) algorithm analysed feature contributions within the optimal model.
Results:
This study included 212 lung cancer patients who were treated with PD-1 inhibitors. The categorical boosting (CatBoost) model demonstrated the best overall predictive performance among the 10 models, achieving an AUC of 0.961 in the training set and 0.871 in the validation set. SHAP results indicated that patients with low tumour-node-metastasis (TNM) staging, high haemoglobin levels, and a history of surgery were more likely to achieve objective response to lung cancer treatment. Additionally, Kaplan-Meier stratification showed that patients with a high probability of response exhibited significantly longer progression-free survival.
Conclusions:
This study successfully developed and validated a CatBoost model to predict the objective response of lung cancer patients receiving PD-1 inhibitor therapy. The model demonstrates high efficacy in predicting short-term objective response, providing a robust tool for clinical evaluation. The accompanying web-based risk platform serves as a hypothesis-generating and decision-supportive tool, holding preliminary clinical potential to support individualised screening and risk stratification in lung cancer patients receiving PD-1 inhibitors.