Inflammatory marker-driven deep learning model for postoperative gastric cancer prognosis

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning

Background:

• Accurate prognostic prediction is crucial for optimizing postoperative management and survival in gastric cancer patients.
• Traditional clinicopathological indicators inadequately capture systemic inflammatory and immunonutritional status, which significantly impact tumor progression and recovery.
• Systemic inflammatory markers like Neutrophil-to-Lymphocyte Ratio (NLR) and Platelet-to-Lymphocyte Ratio (PLR) show prognostic value, but their complex interactions challenge traditional statistical methods.

Purpose of the Study:

• To develop a novel deep learning framework for enhanced postoperative prognostic prediction in gastric cancer.
• To integrate clinical, demographic, and systemic inflammatory variables for more accurate outcome evaluation.
• To overcome the limitations of traditional statistical methods in analyzing complex, nonlinear interactions among prognostic factors.

Main Methods:

• A hybrid deep learning framework combining Gradient-Boosted Decision Tree (GBDT), Tree-Driven Encoder (TDE), and 1D Convolutional Neural Network (1D-CNN) was proposed.
• The GBDT module identified dependencies among variables, TDE created unified binary embeddings, and 1D-CNN learned high-level feature representations.
• Model performance was assessed via cross-validation and compared against traditional machine learning and advanced deep learning models.

Main Results:

• The proposed hybrid deep learning framework demonstrated superior performance in predicting postoperative prognosis compared to traditional and general deep learning models.
• The model effectively captured nonlinear and hierarchical relationships, achieving high predictive accuracy, robustness, and generalization, especially for high-risk patients.
• Consistent performance across multiple experimental conditions confirmed the model's reproducibility and reliability.

Conclusions:

• A data-driven, interpretable deep learning framework was successfully developed for gastric cancer postoperative prognostic prediction.
• The integrated approach provides a comprehensive understanding of inflammation, nutrition, and tumor biology interactions, supporting personalized treatment.
• Future work includes external validation, real-time clinical application, and enhancing model explainability for broader clinical adoption.