Research progress and challenges of multimodal deep learning models for prognosis of gastric cancer Authors: Yang Bo, Cai Hongyi, Li Hao Published: 2026 ABSTRACT AbstractIn the field of oncology research, patient data typically encompasses diverse multimodal characteristics, including age, survival status, radiological imaging features, histopathological features, and genomic information. The integrated analysis of these multidimensional data holds significant scientific value for in-depth understanding of tumor biology, development of precise prognosis models, and identification of potential therapeutic targets. Current research and literature indicate that deep learning-based multimodal data fusion techniques can uncover underlying biological mechanisms behind individual patient differences through comprehensive analysis of integrated multi-dimensional data, thereby substantially improving the accuracy of artificial intelligence (AI) in prognosis prediction. This review examines the research progress of multimodal data integration techniques in gastric cancer prognosis prediction, while conducting an in-depth discussion on innovative applications of interpretable deep learning models along with the challenges and issues they face.Keywords: Gastric cancer prognosis, Artificial intelligence, Deep learning, Multimodal fusion, Interpretability, Precision medicine, Clinical decision support Introduction Gastric cancer, as one of the leading causes of cancer-related deaths worldwide [1], is a highly complex disease involving micro-to macro-level changes in its development. These intricate processes are challenging to fully capture through single-source data. Traditional biomarker discovery relies heavily on manual inspection, which is not only time-consuming but also prone to subjectivity and poor reproducibility. More importantly, single-modality biomarkers often fail to maximize their clinical potential.These differences are not only due to the intrinsic molecular diversity of tumors but are also influenced by various macro- and microenvironmental factors. For example, the specific anatomical location of the tumor within the stomach (such as the cardia, body, or pylorus) may affect the thoroughness of surgical resection and the extent of lymph node dissection. At the same time, the dynamically evolving tumor microenvironment (TME) of gastric cancer—including the state of immune cell infiltration, changes in stromal components, and alterations in oxidative stress levels—also significantly impacts treatment response and prognosis [2, 3].Therefore, integrating multimodal data is crucial for comprehensively understanding cancer characteristics and improving diagnostic and prognostic accuracy. Recent advances in artificial intelligence, especially multimodal deep learning technology, have brought new research breakthroughs for gastric cancer prognosis prediction [4]. Traditional Cox models are inadequate for accurate outcome prediction due to oversimplification [5], and the TNM staging system has limited accuracy in predicting gastric cancer-specific survival rates. Deep learning models integrating multimodal data such as pathological whole-slide images, genomic information, and clinical data have shown significant advantages, providing a new research direction for gastric cancer prognosis modeling [6]. However, the existing research system still faces numerous limitations and challenges in addressing data heterogeneity and integration, model interpretability, and clinical translation. This paper elaborates on these research aspects from the following perspectives: (1) Deep learning drives paradigm change; (2) Core methodologies in tumor AI research; (3)Research Status of Multimodal Gastric Cancer Prognostic Models; (4) Multimodal Dataset and Standardized Evaluation System; (5) Multimodal Data Integration Method; (6) Core Challenges of Multimodal Models in Gastric Cancer; (7) Clinical Outlook and Implementation Path of Multimodal Models in Gastric Cancer.A literature search was performed in PubMed, Web of Science, and IEEE Xplore for articles published up to December 2025, using keywords including “gastric cancer”, “multimodal deep learning”, “prognosis”, “histopathology”, and “radiomics”. Additional relevant studies were identified from reference lists. Key findings and challenges are summarized below Fig. 1. Conceptual diagram of a deep learning-based multimodal prediction model for gastric cancer Deep learning drives paradigm change Deep learning is revolutionizing gastric cancer prognosis assessment by transcending traditional empirical frameworks reliant on limited clinical indicators. With the rapid advancement of deep learning technology, multimodal interpretable models have emerged as a game-changer. These innovative models integrate multidimensional data from imaging, pathology, and genomics to enable cross-dimensional correlation analysis across anatomical, pathological, molecular, and clinical features. By leveraging hierarchical feature learning and adaptive weighting mechanisms, they upgrade static indicator analysis to dynamic individualized risk stratification, constructing personalized prognostic models for precise risk subpopulation identification. Furthermore, visualization techniques address interpretability challenges, ultimately establishing a more comprehensive and accurate medical analysis framework. Bottleneck of single-mode prediction model Single-modality models are typically constructed using data from a single source (such as clinical pathological features or molecular markers), making it challenging to fully capture the high heterogeneity of gastric cancer. Taking histopathological image analysis as an example, the study by Guo et al. [7] employed an automated tumor detector and a MSI detection model trained on the TCGA-CRC-DX dataset for prognosis evaluation. The results showed a patient-level AUC of 0.84 (95% CI 0.72–0.92) with C-index values ranging between 0.68 and 0.71, directly highlighting the performance limitations of single-modality models in complex disease assessments. Moreover, the persistent “black box” nature of deep learning models significantly restricts their clinical application [8]. Single-modality models demonstrate limited effectiveness in complex clinical scenarios, particularly when integrating multi-source information. For instance, in predicting short-term and long-term survival rates for pancreatic ductal adenocarcinoma (PDAC) [9], single-modality models underperformed compared to multimodal AI models. Similarly, in chronic neurological disorders research [10], single-modality functional magnetic resonance imaging (fMRI) models showed lower sensitivity than multimodal counterparts. Existing single-modality models are often designed for specific cancer types, lacking cross-disease adaptability. Studies indicate that different cancer types require distinct feature patterns (e.g., deep genetic variations or minor variations), which single-modality architectures struggle to accommodate [11]. Critical comparison of representative multimodal models See Table 1. Table 1Comparison of representative multimodal deep learning models for gastric cancer prognosisStudy (Ref.)Data modalitiesModel architectureKey performance metrics (95% CI)AdvantagesLimitationsZeng et al. [5]Clinical data (SEER cohort)DeepSurv, CNN1/3/5/10-year survival AUC: 0.825–0.871; C-index: 0.81 (0.78–0.84)Outperforms traditional CoxPH (C-index: 0.755) and RSF (0.766); Retrospective validation on large cohortOnly uses clinical data (no imaging/genomics); Lack of external prospective validationLi et al. [19]H&E-stained pathology images, clinical data, RNA-seqResNet-50 + XGBoostTMB prediction AUC: 0.971 (0.943–0.999); C-index for survival: 0.83 (0.79–0.87)Integrates histopathology and multi-omics; High TMB prediction accuracyRelies on high-quality RNA-seq data (limited availability); No cross-institutional validationWei et al. [20]Pathology images, clinical data, genomic dataMultiDeepCox-SC (multi-task deep learning)3-year survival AUC: 0.85 (0.81–0.89); C-index: 0.82 (0.78–0.86)Multi-task learning (prognosis + molecular subtype prediction); Robust to partial data missingComplex training process; High computational resource requirementsDong et al. (MuMo) [22]CT images, pathology sections, structured reports, clinical dataTransformer-based cross-modal fusionAnti-HER2 therapy response AUC: 0.821 (0.77–0.87); Combined immunotherapy AUC: 0.914 (0.88–0.95)Focuses on treatment response prediction (clinical utility); Captures cross-modal correlations via TransformerHER2-positive subset only (limited generalizability); No cost-effectiveness analysisRichard J. Chen et al. [25]H&E WSI, genomics (TCGA-14 cancers including gastric)AMIL + SNN + multimodal fusion layerC-index (Pan-cancer comprehensive ): 0.644 (0.61–0.68); Improved by 11.4% vs. unimodal pathologyPan-cancer adaptability; Reveals “morphological-molecular” synergyLower C-index than single-cancer models; Batch effects in TCGA WSI data Comparison of representative multimodal deep learning models for gastric cancer prognosis Evidence-based guidance for fusion method selection See Table 2. Table 2Guidance for multimodal fusion method selectionFusion typeCore mechanismSuitable scenariosAdvantagesRecommended tools/studiesEarly fusionInput-level concatenation (vector/Hadamard product) [31–33]High data consistency (e.g., paired pathology-genomics); Small sample sizeCaptures fine-grained cross-modal correlations; Low computational costBilinear pooling [33]; Le et al. [34] (prostate cancer classification)Intermediate fusionDynamic weight allocation via attention (e.g., multimodal Transformer) [38]Variable data importance across samples (e.g., early-stage: imaging-focused; late-stage: genomics-focused)Adaptive to sample heterogeneity; Balances modal contributionsCross-attention Transformer [38]; TCGA survival analysis (C-index: 0.78)Late fusionAggregation of unimodal predictions (logistic regression/MLP) [25]High data heterogeneity (e.g., mismatched imaging-genomics); Partial modal missingRobust to noisy unimodal data; Easy to implement/modifyAMIL + SNN [25]; Improved C-index by 6.8% vs. unimodal in gastric cancer Guidance for multimodal fusion method selection Core methods in tumor AI research AI technology demonstrates tremendous potential in processing multimodal medical data, with its core methodologies categorized into three main types: supervised learning, weakly supervised learning, and unsupervised learning. Each approach has distinct application scenarios and advantages. Supervised learning methods Manual Feature Method: The process begins with extracting predefined features from datasets (such as cell shape or size), followed by training using traditional machine learning models like Random Forest (RF) and Support Vector Machines (SVM) [12, 39]. While these models offer advantages such as simple architecture, low computational costs, and high interpretability, the feature extraction process is time-consuming and may introduce human biases into the model. Deep Learning Methods: Represented by convolutional neural networks (CNNs), these models directly process raw data or image blocks to automatically learn predictive features. By alternately applying convolutional layers, pooling layers, and nonlinear activation layers, CNNs can extract rich feature representations from raw data, offering advantages such as low preprocessing costs, high flexibility, and superior performance. However, they rely on pixel-level annotations that are not only time-consuming but also susceptible to variations in human judgment and human biases, while also resulting in poor model interpretability. Weakly supervised learning methods Due to the impracticality of pixel-by-pixel or block-by-block labeling required by strong supervised learning in clinical workflows, coupled with its high implementation costs, weakly supervised learning methods have emerged as a viable alternative. These approaches enable model training using patient-level weak labels (such as diagnostic outcomes or survival rates) without manual annotation. Key implementations include: Graph Convolutional Networks (GCNs): Leveraging graph structures to capture tissue spatial configurations and inter-object relationships, these networks update node feature representations by aggregating information from neighboring nodes, making them suitable for tasks requiring extensive spatial context [13]. Multiple Instance Learning (MIL): Treating all image blocks from a single patient as a unified “bag” and training exclusively with patient-level labels [14], this method utilizes attention mechanisms to determine each block’s predictive importance, achieving robust performance without precise annotations [15]. Visual Transformer (ViT): Employing attention mechanisms to consider inter-block correlations and contextual information [16, 17], ViT enhances global feature capture through components like position encoding, self-attention, and multi-head attention, thereby improving model robustness. Unsupervised learning methods Unsupervised learning methods, which do not rely on labels, primarily explore structures, patterns, and subgroups within datasets. These include both self-supervised and fully unsupervised approaches. Self-Supervised Learning: Involves designing pretext tasks (such as predicting missing image elements) to learn universal image features from unlabeled data, then applying these features to supervised tasks. This approach enhances model performance even with limited data [18]. Fully Unsupervised Learning: Encompasses methods like clustering and dimensionality reduction. Clustering techniques partition data into distinct subgroups for identifying cancer subtypes, while dimensionality reduction approaches such as autoencoders capture key features and correlations through low-dimensional representations, facilitating data exploration and feature extraction [21, 46]. Research status of multimodal gastric cancer prognostic models Multimodal deep learning models have made significant progress in gastric cancer prognosis research. By integrating multi-source data such as histopathological images (HE-stained sections), medical imaging (CT, endoscopy), genomic data, and clinical information, these models have significantly improved the accuracy of prognosis prediction and clinical applicability [35–37]. Current research primarily employs architectures like convolutional neural networks (CNNs), deep survival neural networks, and graph neural networks. Notably, the clinical utility of these models extends beyond survival prediction to risk stratification for healthcare resource allocation: high-risk patients receive prioritized treatment and enhanced follow-up (e.g., quarterly CT scans), while low-risk patients undergo reduced monitoring (e.g., annual check-ups), optimizing resource utilization [,22–24, 26, 29, 30]. However, challenges persist in data heterogeneity, model interpretability, and clinical validation, as detailed in Sects. 7 and 8. Multimodal dataset and standardized evaluation system Multimodal data types and typical datasets See Table 3. Table 3Multimodal datasets for gastric cancer prognosisDataset nameData typesURLPopulation representationLimitationsTCGA-STADGenome copy number variation, epigenetics, gene expression, miRNA, clinical data, pathology images https://portal.gdc.cancer.gov/ 443 patients (Estimated 58% White, 19% Asian, 23% unknown) [47]Overrepresentation of White patients; Batch effects in gene expression dataTCIA (Gastric cancer subset)CT, MRI, pathology images https://www.cancerimagingarchive.net/ 20 patients (Independent patients: 20 cases; Total imaging sequences: 99༛Total CT images: 38,660) [52]Limited sample size; No standardized scanning protocolsSEERClinical data (age, stage, survival, treatment) https://seer.cancer.gov/ >100,000 gastric cancer patients (U.S.-based, diverse ethnicity)No imaging/genomic data; Variable follow-up completenessKaplan-Meier plotter (gastric module)mRNA expression, survival data https://kmplot.com/analysis/ >1000 patients (European + Asian cohorts)No clinical covariates (e.g., treatment); mRNA data onlyUCSC XENA (gastric)Pan-cancer gene expression, clinical data http://xena.ucsc.edu/ >1000 patients (TCGA + other cohorts)Heterogeneous data sources; Limited pathological annotations Multimodal datasets for gastric cancer prognosis Standardized evaluation metrics To resolve inconsistent metric reporting, Table 4 standardizes key evaluation indicators with definitions, clinical relevance, and example values from gastric cancer studies. Table 4Standardized evaluation metrics for gastric cancer prognosis modelsMetricDefinitionClinical relevanceExample values (gastric cancer models)AUCArea under ROC curve (TPR vs. FPR)Compares overall model performance; Robust to class imbalanceMuMo model (anti-HER2 response): 0.821 (95% CI 0.77–0.87) [22]C-index (Harrell’s)Proportion of patient pairs with correctly ordered survival timesGold standard for survival models; Accounts for censored dataDeepSurv: 0.81 (95% CI 0.78–0.84) [5]; AMIL + SNN: 0.644 (0.61–0.68) [25]F1-scoreHarmonic mean of precision and recallEvaluates performance on imbalanced subgroups (e.g., high-risk patients)Li et al.’s TMB prediction: 0.92 (0.88–0.96) [19]Brier scoreMean squared error between predicted and actual survival probabilitiesMeasures calibration (prediction confidence vs. actual risk)MultiDeepCox-SC: 0.12 (0.10–0.14) [20] (lower = better)Log-rank test p-valueStatistical significance of survival difference between predicted risk groupsValidates clinical utility of risk stratificationZeng et al.’s model: p < 0.001 [5] Standardized evaluation metrics for gastric cancer prognosis models Data quality and fairness Key data quality issues and mitigation strategies Data Bias & Racial Representation: TCGA-STAD underrepresents Asian and African populations [47], leading to reduced generalizability to non-White cohorts. Mitigation: Establish multi-center, cross-ethnic datasets (e.g., Asia-Pacific Gastric Cancer Cohort [48]) and apply fairness-aware training (e.g., reweighting underrepresented groups [49]). Batch Effects: Genomic/imaging data variations across platforms/scanners (e.g., TCGA gene expression batches) distort features. Mitigation: Use batch correction methods like ComBat [50] (for genomics) or CycleGAN [51] (for imaging). Missing Data: Estimated 15–30% of clinical/genomic data is missing in TCGA-STAD. Mitigation: Multimodal imputation (e.g., EmbraceNet [45], which uses complete modalities to compensate for missing ones) or MICE (Multiple Imputation by Chained Equations) [53]. Privacy Protection: In the field of medical artificial intelligence, the privacy of patient data is subject to strict regulations (such as HIPAA and GDPR), which greatly limits cross-institutional data sharing, thereby hindering the development and validation of high-performance models [54]., To address this challenge, researchers have proposed various privacy protection technologies, mainly including Federated Learning, Differential Privacy (DP), and federated multimodal fusion methods.Differential privacy provides strong privacy protection for individual records in theory by injecting controllable random noise during the model training process [55].Similarly, federated multimodal fusion is also regarded as an effective privacy-preserving paradigm [56]. Prospective data collection guidelines To improve data quality, future datasets should adhere to:1.Standardized protocols (e.g., TCGA imaging guidelines [57]) for scanning, staining, and sequencing.2.Prospective design with predefined endpoints (e.g., 5-year overall survival) and complete follow-up.3.Diverse population enrollment (minimum 30% non-White patients [48]) 0.4.Linked multimodal data (paired pathology-imaging-genomics from the same patient). Multimodal data integration method Data types and preprocessing See Table 5. Table 5Multimodal data types and preprocessingModality typeData sourcesPreprocessing technologiesTypical feature dimensionsMedical imagingAbdominal CT (DICOM)3D U-Net segmentation + PyRadiomics feature extraction + CycleGAN batch correction [51]385 radiomics featuresGenomicsTCGA-STAD cohortWhole exome sequencing + somatic mutation annotation + ComBat batch correction [50]20,300 gene lociClinical dataEMR systemStandardized encoding (SNOMED CT) + MICE missing value imputation [53]56 clinical variablesProteomicsOlink Explore 1536 panelMass spectrometry data normalization + pathway enrichment analysis1,536 protein markersPathomicsTCGA-STAD WSIOpenslide image reading + Masking + Histogram equalization (for color consistency)432 cases, 755 WSI Multimodal data types and preprocessing Explainable technology system See Table 6. Table 6Core explainable methods for gastric cancer multimodal modelsMethodApplication scenarioImplementation toolClinical validationGrad-CAMCT/pathology lesion localizationtorchcam, tf-kerasValidated by pathologists: estimated 89% agreement between Grad-CAM heatmaps and manual annotations [41–44, 58]SHAPFeature contribution analysis (e.g., TNM staging weight)Shap librarySHAP can interpret the prediction results of models such as SVM, RF, and LR, enhancing clinical applicability and supporting individualized treatment decisions [59].LIMESingle-sample prediction explanationLime libraryLIME is used to explain the predictions of machine learning models, in order to enhance doctors’ trust in model outputs and improve the effectiveness of decision support in a medical context [60]. Core explainable methods for gastric cancer multimodal models Core challenges of multimodal models in gastric cancer Data-level: gastric cancer-specific data dilemmas Gastric Cancer Multimodal Model Related TextHeterogeneity and Annotation Difficulties: Gastric cancer exhibits significant morphological variations across subtypes (intestinal/diffuse type). Annotation of H&E-stained WSIs requires pathologists to distinguish subtype-specific features such as “signet ring cells” and “tubular structures,” with the cost of a single WSI annotation being substantial and generally higher than that for lung cancer. Moreover, inter-center annotation consistency for gastric cancer is moderate, which is lower than the typically high consistency reported for breast cancer. For CT images, the subjective judgment of tumor invasion depth in gastric cancer staging (T1–T4) and cross-device (e.g., 16‑slice vs. 64‑slice) differences lead to a noticeable drop in the model’s inter-center C‑index [63]. High Missing Rate and Poor Clinical Adaptability: An estimated 40%–60% of gastric cancer samples have missing modalities in clinical settings—early-stage patients often omit genomic testing due to the “minimally invasive priority” principle, while advanced patients lack post-operative follow-up CT due to poor performance status [64]. Additionally, an estimated 80% of primary hospitals have unstructured gastric cancer EMRs, making it impossible to directly convert key pathological descriptions (e.g., “poor differentiation with mucin lakes”) into model features [65]. Clinical translation: implementation barriers in gastric cancer practice Multimodal models face challenges of extended regulatory and trial cycles during clinical translation. These models need to be validated across the entire chain from prognosis prediction → treatment decision → survival benefit, whereas prospective clinical trials usually require estimated 3–5 years of follow-up. For example, a clinical trial with the estimated 3-year disease-free survival rate after gastric cancer surgery as the primary endpoint has a follow-up period significantly longer than the estimated 2-year follow-up design commonly seen in lung cancer studies. In addition, gastric cancer is highly heterogeneous, including molecular and histological subtype differences such as Lauren classification (intestinal vs. diffuse type) and HER2 status, which requires multimodal models to maintain robust predictive performance across these subgroups as much as possible (e.g., C-index in the range of estimated 0.726–0.797), thereby markedly increasing the complexity of trial design and validation [61–68]. Cost‑Workflow Conflicts: Building a 1,000‑patient gastric cancer multimodal dataset involves substantial cost (including subtype‑specific annotation and multi‑center ethical review), which is unaffordable for small and medium‑sized hospitals [69]. Current models require manual entry of clinical information (e.g., HER2 test results, chemotherapy regimens), which is time‑consuming in specific clinical settings—significantly extending the traditional gastric cancer management workflow [22]. Physician Distrust and Inadequate Training: A considerable proportion of gastric cancer clinicians question model outputs. For example, when a diffuse‑type gastric cancer patient with favorable clinical indicators is predicted as high‑risk, clinicians reject the AI result because they “cannot understand why the model prioritizes the ‘tumor invasion of serosa’ feature.” Only a small minority of gastric cancer clinicians have received training on “SHAP plot interpretation,” limiting their ability to link model features to clinical decisions [57, 70]. Ethics and fairness: gastric cancer population-specific risks Data Privacy and Ethnic Bias: Most multi-center gastric cancer datasets originate from Asian populations (≥ estimated 70%) [28]. When using federated learning, compliance with both HIPAA and China’s Personal Information Protection Law is required, and cross-regional parameter transmission demands additional encryption (increasing computational costs by estimated 20%–30%). Models tend to have prediction biases for Caucasian gastric cancer patients (estimated 8%–10% C-index difference) due to their underrepresentation (only estimated 15%) in training data [49]. Interpretability Compliance Pressure: The FDA mandates “subtype-specific interpretability reports” for gastric cancer multimodal SaMD. For example, for HER2-positive patients, the fusion logic between “CT enhancement degree” and “HER2 expression level” must be clarified. However, current models only output overall risk scores and cannot decompose feature contributions specific to subtypes [62]. Clinical outlook and implementation path of multimodal models in gastric cancer Emerging technologies: targeting unmet needs in gastric cancer management Specialized foundation models (Foundation Models, FM), as an emerging technology, show great potential in addressing unmet needs in the diagnosis and treatment of gastric cancer. These models are mainly based on the Transformer architecture and combine self-supervised pretraining strategies, encompassing various forms such as large language models (LLM), vision foundation models (VFM), and multimodal large language models (Multimodal LLM). They are widely applied in multiple key areas, including clinical narrative structuring, medical record information extraction, endoscopy/imaging/pathology image analysis, diagnostic support, treatment response prediction, and prognosis evaluation [70–72]. Clinical impact: advancing precision gastric cancer care Risk Stratification and Precision Treatment: The model can stratify Stage II gastric cancer into “high-risk (estimated 5-year survival < 60%, SOX chemotherapy recommended)” and “low-risk (estimated 5-year survival > 75%, observation recommended),” reducing overtreatment by an estimated 25%–30% in specific research cohorts [5]. For HER2-positive patients, the MuMo model predicts response to anti-HER2 therapy (AUC = estimated 0.914), avoiding ineffective treatment [22]. Resource Optimization: Tailored follow-up strategies (“intensive follow-up for high-risk patients [CT every estimated 3 months] + relaxed follow-up for low-risk patients [CT every estimated 6 months]”) reduce annual gastric cancer follow-up costs by an estimated 18%–22% in specific clinical settings [40]. Key limitations impeding the generalizability of model metrics Although multimodal deep learning models have shown promising performance in gastric cancer prognosis prediction, the reported metrics (e.g., C-index, AUC) are mostly derived from retrospective single-center or public dataset studies. Their clinical generalizability is closely linked to the limitations discussed earlier, primarily constrained by four core factors: Cohort Bias: Differences in abdominal CT scanners, non-standardized pathological staining protocols, and genomic detection platforms across research centers, coupled with racial representation bias in public datasets (e.g., overrepresentation of White patients in TCGA-STAD), make models difficult to adapt to the Asian population with a high incidence of gastric cancer, directly reducing cross-center and cross-population applicability [47, 63]. Cohort Bias: Differences in abdominal CT scanners, non-standardized pathological staining protocols, and genomic detection platforms across research centers, coupled with racial representation bias in public datasets (e.g., overrepresentation of White patients in TCGA-STAD), make models difficult to adapt to the Asian population with a high incidence of gastric cancer, directly reducing cross-center and cross-population applicability [47, 63]. Non-random Missing Modalities: Approximately estimated 40%–60% of gastric cancer clinical samples have modality missing, which is associated with disease stage (early-stage patients lack genomic data, advanced patients lack follow-up imaging). This non-random pattern makes training datasets fail to reflect real clinical characteristics, leading to poor reproducibility of model metrics in real-world scenarios. Non-random Missing Modalities: Approximately estimated 40%–60% of gastric cancer clinical samples have modality missing, which is associated with disease stage (early-stage patients lack genomic data, advanced patients lack follow-up imaging). This non-random pattern makes training datasets fail to reflect real clinical characteristics, leading to poor reproducibility of model metrics in real-world scenarios. Heterogeneity of Outcome Definitions: Inconsistent survival outcome definitions (overall survival [OS], disease-free survival [DFS], progression-free survival [PFS]) across studies lack a unified standard, hindering horizontal comparison of model metrics and increasing clinical translation difficulty [5, 20, 22, 28, 67]. Heterogeneity of Outcome Definitions: Inconsistent survival outcome definitions (overall survival [OS], disease-free survival [DFS], progression-free survival [PFS]) across studies lack a unified standard, hindering horizontal comparison of model metrics and increasing clinical translation difficulty [5, 20, 22, 28, 67]. Risk of Data Leakage: In multimodal gastric cancer research (especially pathomics analysis of H&E-stained WSIs), unclear splitting of pathological slices/cases between training and validation sets (e.g., multiple slices from the same patient scattered in different sets) leads the model to pre-learn patient-specific features, overestimating core metrics such as C-index and AUC [13, 31]. Risk of Data Leakage: In multimodal gastric cancer research (especially pathomics analysis of H&E-stained WSIs), unclear splitting of pathological slices/cases between training and validation sets (e.g., multiple slices from the same patient scattered in different sets) leads the model to pre-learn patient-specific features, overestimating core metrics such as C-index and AUC [13, 31]. In summary, these four factors are the key links between existing model limitations and poor metric generalizability. Future research should focus on establishing multi-center standardized gastric cancer datasets, developing non-random missing data imputation algorithms, formulating unified outcome definition standards, and strictly implementing data splitting rules to enhance the generalizability of model metrics and promote clinical translation. Author contributions B.Y. and H.C. designed the study, collected and analyzed the data, and drafted the main manuscript text. H.L. prepared the figures and tables, and revised the manuscript for critical intellectual content. All authors reviewed and approved the final version of the manuscript for submission. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. Data availability No datasets were generated or analysed during the current study. Declarations Ethics approval and consent to participate Not applicable. This study is a narrative review of published literature and does not involve any clinical research on human or animal subjects. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s Note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References Associated Data Data Availability Statement No datasets were generated or analysed during the current study.