Sintilimab plus chemotherapy has proven effective as a combination immunotherapy for patients with advanced gastric and gastroesophageal junction adenocarcinoma (GC/GEJC). A multi-center study conducted in China revealed a median progression-free survival (PFS) of 7.1 months. However, the prediction of response duration to this immunotherapy has not been thoroughly investigated. Additionally, the potential of baseline laboratory features in predicting PFS remains largely unexplored. Therefore, we developed an interpretable machine learning (ML) framework, iPFS-SC, aimed at predicting PFS using baseline (pre-treatment) laboratory features and providing interpretations of the predictions.
A cohort of 146 patients with advanced GC/GEJC, along with their baseline laboratory features, was included in the iPFS-SC framework. Through a forward feature selection process, predictive baseline features were identified, and four ML algorithms were developed to categorize PFS duration based on a threshold of 7.1 months. Furthermore, we employed explainable artificial intelligence (XAI) methodologies to elucidate the relationship between features and model predictions.
The findings demonstrated that LightGBM achieved an accuracy of 0.70 in predicting PFS for advanced GC/GEJC patients. Furthermore, an F1-score of 0.77 was attained for identifying patients with PFS durations shorter than 7.1 months. Through the feature selection process, we identified 11 predictive features. Additionally, our framework facilitated the discovery of relationships between laboratory features and PFS.
A ML-based framework was developed to predict Sintilimab plus chemotherapy response duration with high accuracy. The suggested predictive features are easily accessible through routine laboratory tests. Furthermore, XAI techniques offer comprehensive explanations, both at the global and individual level, regarding PFS predictions. This framework enables patients to better understand their treatment plans, while clinicians can customize therapeutic approaches based on the explanations provided by the model.