Predicting Intratumoral Heterogeneity and Stratifying Prognostic Risk in Gastric Cancer Using a Histology-Based...

Intratumoral heterogeneity (ITH) is a fundamental driver of clonal evolution and therapeutic resistance in gastric cancer (GC). However, the clinical assessment of ITH remains limited by the high cost and technical complexity of multiregion and single-cell sequencing. This study aimed to develop a pathomics-deep learning (DL) fusion model for estimating ITH directly from routine hematoxylin and eosin-stained whole-slide images in GC. We retrospectively collected 893 whole-slide images from 773 patients in 3 independent cohorts. ITH was quantified using 8 algorithms, and the optimal prognostic indicator was identified by Cox regression analysis. Pathomics features were extracted using CellProfiler, whereas DL features were generated using ResNet50 combined with 2 multiple instance learning pipelines. A pathology-driven stacking ensemble model integrating pathomics and DL features was then constructed. A total of 239, 103, 135, and 30 patients were included in the training, internal validation, and 2 external test cohorts, respectively. The mutant-allele tumor heterogeneity (MATH) score was identified as an independent prognostic factor for overall survival (hazard ratio, 1.840; 95% CI, 1.144-2.957; P =.012). Biological relevance analysis showed that high-MATH tumors were characterized by increased chromosomal instability and an immunosuppressive microenvironment, whereas low-MATH tumors were associated with immune-active phenotypes. The pathology-driven stacking ensemble model showed favorable performance in predicting MATH-defined ITH status, with areas under the curve of 0.852 to 0.956 across the training, validation, and test cohorts. Moreover, the model-derived ITH risk score was an independent prognostic factor and effectively stratified patients into high- and low-risk groups with significantly distinct overall survival outcomes (all P <.05). Further interpretability analysis of the DL component using Gradient-weighted Class Activation Mapping (Grad-CAM) showed that the model primarily attended to regions characterized by nuclear atypia and immune-cell infiltration. In conclusion, we developed a robust and interpretable hematoxylin and eosin-based model for predicting MATH-defined ITH and supporting prognostic stratification in patients with GC. This artificial intelligence-driven approach may provide a cost-effective and scalable tool to support precision oncology in clinical practice.