Efficacy prediction of first-line immunochemotherapy for patients with advanced gastric cancer:Retrospective,Machine Learning study (Preprint)

Abstract

BACKGROUND Immunochemotherapy has brought new hope for the first-line treatment of advanced gastric cancer(GC) patients; however, there is still a lack of simple and effective models to predict the efficacy of immunochemotherapy in this setting. OBJECTIVE This study aimed to develop prognostic models and chose the better one to assess the benefit of first-line immunochemotherapy in patients with advanced GC. METHODS To predict the efficacy of first-line immunochemotherapy in advanced GC patients, we retrospectively collected clinical data at The First Affiliated Hospital of Nanjing Medical University between January 2018 and October 2023. The dataset was split into training (70%) and validation (30%) sets. Additionally, a temporal validation cohort between November 2023 and September 2024 was used to further assess model performance over time. Univariate and multivariate Cox regression, least absolute shrinkage and selection operator (LASSO) regression, and clinical experience were used to select significant features associated with the response of first-line immunochemotherapy in advanced GC patients. Two survival models, LASSO-Cox and Random Survival Forest (RSF) were developed using the training set. Model performance was evaluated on the validation set using an area under the curve (AUC), continuous concordance index, time-dependent receiver operating characteristic (ROC) curves, and calibration plots. Additionally, decision curve analysis(DCA) was performed to evaluate clinical utility, and Kaplan-Meier curves were used to compare survival outcomes. RESULTS In this study, we compared the performance of the RSF model and the LASSO-Cox model in predicting progression-free survival (PFS). The ROC curves showed that the RSF model significantly outperformed the LASSO-Cox model in the training and validation cohorts, with higher AUCs at 6, 12, and 18 months. Additionally, in the ROC curves for the temporal validation cohort, the RSF model demonstrated superior discriminatory ability compared to the LASSO-Cox model. Calibration curves indicated that the RSF model demonstrated good discrimination for PFS predictions. Continuous AUCs and C-index values consistently showed the RSF model's robust performance across different time points. In terms of clinical application, DCA curves demonstrated that the RSF model yielded greater net benefits compared to the LASSO-Cox model. We calculated risk scores for each patient using the RSF model, categorizing them into high-risk (risk score ≥ 55.42) and low-risk (risk score < 55.42) groups. Kaplan-Meier analysis revealed that the low-risk group had a significantly higher progression-free survival rate than the high-risk group, further highlighting the RSF model's importance in patient risk stratification. In summary, the RSF model exhibited superior accuracy and clinical utility in predicting PFS, particularly for long-term predictions compared to the LASSO-Cox model. CONCLUSIONS This study retrospectively collected clinical data from advanced GC patients undergoing first-line immunochemotherapy and established LASSO-Cox and RSF models. By comparing the discrimination, calibration, and clinical utility of the two models across the training set, internal validation set, and temporal validation cohort, we found that the RSF model consistently outperformed the LASSO-Cox model. This highlights the RSF model's superior predictive accuracy and its potential to offer valuable insights for improving clinical decision-making in advanced GC.