JMIR Publications

ABSTRACT

Background:

Immune checkpoint inhibitors represent an effective therapeutic approach for advanced gastric cancer. Their efficacy largely depends on the status of tumor biomarkers including HER2, PD-L1 (CPS ≥1), and MSI-H. To non-invasively evaluate these biomarkers, researchers have developed radiomic models for individual biomarker prediction. However, in clinical practice, holistic prediction of these biomarkers as an integrated system is more efficient. Currently, the feasibility of implementing radiomics-based comprehensive biomarker prediction remains unclear, requiring further investigation.

Objective:

This study aimed to develop a radiomics-based predictive model using multi-phase CT images to holistically evaluate HER2, PD-L1, and MSI-H status in patients with gastric cancer.

Methods:

A retrospective analysis was conducted on 461 gastric cancer patients who underwent radical gastrectomy between 2019 and 2022. Clinical data, contrast-enhanced CT images (arterial phase [AP] and portal venous phase [PP]), and pathological results were collected. Patients were categorized into two groups: the PD-1 inhibitor panel positive group, comprising patients with HER2 overexpression, PD-L1 positive, or MSI-H status, and the negative group, comprising patients without HER2 amplification, PD-L1 negative, or MSI-L/MSS condition. Radiomic features (including first-order statistics, shape features, and wavelet-derived textures) were extracted from both AP and PP images, yielding 1,834 features per phase (total 3,668 features). Least Absolute Shrinkage and Selection Operator (LASSO) regression was applied to select key features. Three models were constructed using the XGBoost algorithm: AP-only (8 features), PP-only (22 features), and a fused model combining AP and PP features (20 features). Model performance was evaluated using area under the curve (AUC), sensitivity, specificity, and decision curve analysis (DCA).

Results:

Of the 461 patients, 147 patient (31.9%) were classified into the panel positive group. The clinical features were similar between the two groups. The fused model demonstrated superior performance in the test cohort (AUC 0.82, 95% CI 0.68–0.95), significantly outperforming AP-only (AUC 0.61, 95% CI 0.47-0.74) and PP-only models (AUC 0.70, 95% CI 0.49-0.91). Sensitivity and specificity for the AP-only, PP-only, and the fused model were (0.33, 0.85), (0.50, 0.86), and(0.60, 0.83), respectively. DCA confirmed that the fused model provided higher clinical net benefit across threshold probabilities.

Conclusions:

The construction of integrated biomarker prediction models through radiomics demonstrates technical feasibility, offering a promising methodology for comprehensive tumor characterization.