# A Machine Learning Model Based on Clinical Factors to Predict the Efficacy of First-Line Immunochemotherapy for Patients With Advanced Gastric Cancer: Retrospective Study

## TL;DR

This study develops a machine learning model to predict which advanced gastric cancer patients will benefit most from immunochemotherapy, using clinical data to guide personalized treatment.

## Contribution

The study introduces a novel machine learning approach using clinical factors to predict immunochemotherapy efficacy in advanced gastric cancer.

## Key findings

- The random survival forest model outperformed other models in predicting progression-free survival.
- Key factors like age, liver metastasis, and immune cell proportions were identified as important predictors.
- The model successfully stratified patients into high- and low-risk groups with distinct survival outcomes.

## Abstract

The development of immunotherapy has provided new hope for patients with advanced gastric cancer (AGC). However, due to the high heterogeneity of the disease, the efficacy of first-line immunochemotherapy varies among patients. There is still a lack of simple and effective models to predict the efficacy of immunochemotherapy in this setting.

This study aimed to identify critical factors and develop predictive models to evaluate the efficacy of first-line immunochemotherapy in patients with AGC using clinically available data. The goal was to offer evidence-based guidance for clinical practice and enable personalized treatment strategies.

To evaluate the effectiveness of first-line immunochemotherapy in AGC, we retrospectively collected clinical data from The First Affiliated Hospital of Nanjing Medical University between January 2018 and October 2023. The data collected were divided into a training set (168/240, 70%) and an internal validation set (72/240, 30%). Additionally, a temporal validation cohort of 76 patients recruited from November 2023 to September 2024 was assembled to further evaluate the predictive performance of the models. We used univariate and multivariate Cox regression analyses, along with the least absolute shrinkage and selection operator (LASSO) regression, and integrated clinical expertise to identify key predictors of treatment efficacy and to construct the LASSO-Cox model. We developed 4 models (LASSO-Cox, random survival forest [RSF], extreme gradient boosting, and survival support vector machine) and evaluated their performance using the C-index, area under the curve (AUC), calibration curves, and decision curve analysis. The optimal model was interpreted using Shapley additive explanations, and its risk scores were used to stratify patients for Kaplan-Meier survival analysis.

Among the 4 prognostic models developed in this study, the RSF model demonstrated superior predictive accuracy and discrimination for progression-free survival, as evidenced by its higher AUC, concordance index, continuous AUC curves, and calibration curves compared with the other 3 models. Additionally, decision curve analysis showed that the RSF model offered greater net clinical benefit. The Shapley additive explanations results identified that age, histological subtype, the proportion of CD19+ B cells, CD16+CD56+ natural killer cells, and the presence of liver metastasis were key prognostic factors influencing patient outcomes. Patients in the low-risk group, as determined by the RSF model’s risk score, exhibited a significantly higher progression-free survival rate than those in the high-risk group, further validating the value of the RSF model for risk stratification.

This study is the first to use machine learning algorithms to develop a predictive model for the efficacy of first-line immunochemotherapy in AGC, and to identify key predictors of treatment outcome. The results indicate that the RSF model not only enables precise stratification of patients likely to benefit but, more importantly, provides quantifiable decision support for individualized clinical strategies, underscoring its potential value in clinical decision-making.