Bridging artificial intelligence and clinical decision-making in gastric cancer surgery

Abstract

Postoperative survival among gastric cancer patients after potentially curative surgery can vary significantly. Therefore, a valid and individualized predictor of postoperative survival in gastric cancer patients who have undergone potentially curative resection is urgently needed. This article highlights the use of interpretable machine learning for individualized risk assessment in gastric cancer patients and provides initial insights into future directions, including multicenter studies incorporating deeper biological insights.

Key Words: Gastric cancer; Postoperative survival; Prognostic model; Machine learning; Surgical oncology

Core Tip: Artificial intelligence has the potential to transform gastric cancer surgery in multiple ways, including improving risk prediction for perioperative complications and enhancing surgical planning, treatment selection, and postoperative monitoring through the use of large-scale clinical data. Beyond achieving high predictive accuracy, artificial intelligence models must also be interpretable, transparent and clinically applicable. As these technologies continue to evolve, they are expected to become increasingly valuable tools in surgical practice. This article highlights how clinically transparent artificial intelligence models can support individualized risk stratification and inform postoperative decision-making in gastric cancer surgery.

INTRODUCTION

Gastric cancer is one of the leading causes of cancer deaths worldwide. For patients with resectable gastric cancer, curative gastrectomy remains the preferred treatment. Although improvements have been made in surgical methods, perioperative care, and adjuvant therapy, the postoperative survival of patients with gastric cancer has not improved. There is still a large variation in the long-term survival of patients with the same clinical and pathological characteristics and the same tumor-node-metastasis (TNM) stage[1-3]. These discrepancies highlight the limitations of anatomy-based prognostic models in guiding postoperative clinical decision-making.

TNM staging as a form of anatomical classification aids in treatment decisions for the patient. Yet, it does not report tumor size nor does it consider the biological characteristics of the tumor, the patient’s reaction and the surgical aspects influencing survival. During postoperative care, healthcare providers are confronted with various clinical situations that need to be handled with little evidence to inform their decisions. Scientists are developing new cancer treatment predictive models to aid in these situations. These models will enable doctors to assess the risk to the patient based on various factors. Scientists are also applying artificial intelligence (AI)-powered prognostic models to the surgical management of gastric cancer. Gastric cancer is a highly heterogeneous disease with many variables that need to be managed in the operating room. Outcomes for patients with colorectal and esophageal cancers are generally predictable and correlate with the specific disease stage. In contrast, gastric cancer is generally an unpredictable and challenging disease to treat due to unpredictable tumor biology and patient-specific factors, including nutritional status and preoperative inflammation. In addition to the tumor itself, other surgical-related variables can impact a patient’s postoperative outcome, including increased blood loss and complications. With many variables that can affect a patient’s outcome, gastric cancer is an ideal disease to apply machine learning (ML) approaches to develop the most accurate individualized survival estimate beyond that predicted by the classic cancer staging system.

BIOLOGICAL AND SYSTEMIC DETERMINANTS OF POSTOPERATIVE HETEROGENEITY

Tumor-intrinsic biological heterogeneity is the most important factor that contributes to postoperative outcomes of patients who have undergone curative gastrectomy. In addition to clinical stage, the clinical behavior of gastric cancer is influenced by the spatial, molecular and phenotypic heterogeneity within the tumor. The postoperative survival of patients with HER2-positive gastric cancer was found to deteriorate in cases where the degree of HER2 expression within the tumor showed considerable variance[4,5]. Single markers and anatomical models do not accurately portray tumor complexity or host relationship. Studies have found that many of the host responses to tumors and their environment are triggered by tumor-induced surgical inflammation. These responses are often host-specific and can have both short-term and long-term effects on the patient’s surgical outcome. Recent studies have found that early postoperative C-reactive protein levels correlate with surgical outcomes and are elevated in patients with poor oncologic outcomes. Systemic inflammation can suppress antitumor immune responses[6]. Cancer surgery can cause a high degree of acute inflammation, which can subsequently cause an acute decrease in immunity. Furthermore, within the healthcare system, management of cancer patients undergoing surgery can involve several surgical events and clinical scenarios. Research findings have found that postoperative complications of cancer surgery, particularly “infectious complications”, can have a substantial potentially permanent impact on quality of life and lead to a decrease in survival unrelated to tumor stage that persists even after treatment for recurrence[7]. Postoperative complications can cause a persistent inflammatory and immunosuppressive state, leading to uncontrolled tumor growth. Differences in the prognosis of patients who have undergone curative surgery for gastric cancer are primarily due to a combination of factors, including tumor heterogeneity, and the degree of systemic inflammation and perioperative events.

THE NEED FOR INTEGRATIVE AND INTERPRETABLE MACHINE LEARNING

The development of AI-based models for postoperative prediction became possible through recent studies that combined various predictive factors into single prognostic systems. Deep learning and radiomics methods show promise for survival prediction through preoperative computed tomography imaging data evaluation because they reveal tumor characteristics that standard imaging techniques fail to detect[8]. Research studies using molecular and immune methods have discovered gastric cancer subgroups that have different survival patterns and reactions to treatments[5]. ML models demonstrate effective results for gastric cancer treatment, but they have not been widely adopted in medical practice. Current models present multiple methodological problems because they do not properly address missing data points and they produce complex models that are difficult to understand. High-performing algorithms operate as black boxes, leading to lack of trust, transparency and accountability, preventing their use in standard medical care. The study by Lü et al[9] forms the foundation for this current study, and it presents a timely and well-designed investigation. The authors address the common challenge of missing clinical data in surgical cohort studies by introducing a novel hybrid imputation approach that combines distance-based imputation, matrix factorization, and Gower similarity. Their subsequent analyses demonstrate that model performance is highly dependent on imputation quality, highlighting the importance of accurate data handling for reliable prognostic predictions. By integrating rigorous feature selection with SHapley Additive exPlanations (SHAP) analysis, the model achieves both strong predictive power and clinical interpretability. While the methodological framework is a major strength, further evaluation is needed to assess potential biases introduced by the imputation strategies and to determine whether the model complexity is appropriate for the size and characteristics of retrospective surgical datasets. Advanced hybrid imputation methods that use sophisticated techniques to estimate missing values may inadvertently introduce artificial relationships between variables or reduce the underlying complexity of the data, potentially leading to improved performance during internal validation without reflecting true generalizability. These concerns underscore the importance of external validation, prospective cohort studies, and sensitivity analyses when developing ML-based prognostic models for gastric cancer surgery. Such evaluations are essential to assess model robustness, minimize the risk of overfitting, and determine readiness for clinical implementation.

CURRENT LANDSCAPE OF ARTIFICIAL INTELLIGENCE IN GASTRIC CANCER SURGERY

The use of AI in gastrointestinal surgery is rapidly evolving. AI applications now span the entire surgical pathway, including the preoperative, intraoperative and postoperative periods, and are increasingly used in the surgical decision-making process[10]. ML algorithms have demonstrated improved predictive performance for patient- and treatment-specific outcomes, including overall and disease-specific survival, postoperative complications, and cancer recurrence, through the analysis of large and complex datasets[11-14].

The rapid development of high-performance predictive models in clinical and translational medicine has been driven by the growing availability of clinical data, advances in imaging technologies, and increased computing power capable of processing large volumes of high-dimensional information[15,16]. Gastric cancer is characterized by complex tumor biology, dynamic host-tumor interactions, and many perioperative factors, all of which must be considered when developing comprehensive predictive models that integrate both biological and clinical information[17-20].

Multimodal data - including clinical, radiological, intraoperative, and molecular information – can be integrated through ML pipelines that incorporate data preprocessing, feature selection, and model development[21,22]. Explainable AI methods further enhance model interpretability, facilitating risk stratification and personalized clinical decision-making throughout the perioperative continuum. Importantly, AI applications in surgery extend beyond prognostic and time-series prediction models. AI algorithms are increasingly being used for diagnostic imaging, intraoperative guidance, and postoperative monitoring and recovery assessment[23-25]. Figure 1 illustrates a comprehensive AI–driven clinical decision-making framework for gastric cancer surgery, integrating multimodal data sources, ML-based predictive modeling, and explainable outputs to support risk stratification and personalized postoperative management.

Figure 1 Conceptual framework of artificial intelligence–driven clinical decision-making in gastric cancer surgery. The figure was created with assistance from ChatGPT for visual design and language support. However, the scientific content, interpretation, and overall concept of the figure were developed entirely by the authors. CRP: C-reactive protein; NLR: Neutrophil-to-lymphocyte ratio; PET-CT: Positron emission tomography-computed tomography; ML: Machine learning; AUC: Area under the curve; AI: Artificial intelligence; SHAP: SHapley Additive exPlanations.

ADVANCES IN MACHINE LEARNING-BASED PROGNOSTIC MODELING

ML approaches offer several advantages over traditional statistical methods for prognostic modeling, particularly in settings involving large datasets, numerous variables, and complex interactions among features[26,27]. A variety of supervised learning techniques have been applied to predict surgical outcomes in gastrointestinal oncology, including Random Forests, Support Vector Machines, and more recently, Gradient Boosting-based algorithms[28-30].

Deep learning methods can also be applied to the analysis of medical images, including computed tomography (CT) scans and histopathological specimens, through the use of convolutional neural networks (CNNs). These models are capable of identifying complex image features and patterns that may not be detectable through conventional image analysis techniques[31-33].

Radiomics-based approaches further enhance prognostic modeling by extracting high-dimensional quantitative features that capture tumor heterogeneity from imaging data. When integrated with clinical variables, these imaging-derived features can improve the prediction of patient outcomes, including overall survival, and support more accurate risk stratification[34,35].

As summarized in Table 1, these ML approaches offer substantial advantages in predictive performance and the integration of diverse data types. However, they also present important challenges, including limited interpretability, susceptibility to overfitting, and difficulties in achieving robust external validation across independent patient cohorts.

Table 1 Overview of machine learning approaches used in prognostic modeling in gastric cancer surgery.

Method	Application	Strengths	Limitations
Random forest	Survival prediction, complication risk	Handles nonlinear data, robust	Limited interpretability
Support vector machine	Classification of outcomes	High accuracy in small datasets	Sensitive to parameter tuning
Gradient boosting (XGBoost)	Risk prediction models	High predictive performance	Risk of overfitting
Deep learning (CNN)	Imaging analysis (CT, histopathology)	Detects complex patterns	Black-box nature
Radiomics models	Tumor heterogeneity analysis	Integrates imaging features	Reproducibility challenges

XGBoost: Extreme gradient boosting; CNN: Convolutional neural network; CT: Computed tomography.

SURGICAL DATA SCIENCE AND INTRAOPERATIVE ARTIFICIAL INTELLIGENCE APPLICATIONS

Surgical data science has evolved from a predominantly preoperative discipline to one that increasingly supports intraoperative decision-making, reflecting the growing need for real-time guidance during surgery. AI-based computer vision technologies are being used not only for the acquisition and processing of large volumes of surgical data, but also for the automated analysis of anatomical structures, procedural workflows, and instrument activity, thereby improving the accuracy and efficiency of surgical practice[36,37].

Current AI systems are primarily designed to reduce intraoperative errors and enhance surgical performance through the quantitative assessment of technical skills. However, the integration of AI into next-generation robotic surgical platforms has the potential to shift these systems from purely assistive toward semi-autonomous surgical support. Such advances could substantially expand surgical capabilities beyond current state of the art, enabling greater precision and consistency during complex procedures. In addition to the mentioned fields of application, context-aware systems could benefit from the integration of intraoperative AI for feedback and decision support in surgery[38]. Such a system would process intraoperative video and surgical instrument kinematics in real-time, track procedure progress, identify the relevant steps, recognize deviations from standard procedure, and inform the surgeon about potential risks. The systems could be particularly important for oncological procedures, where small differences in technique can have significant impacts on postoperative recovery. Advanced imaging technologies are also expected to play an increasingly important role in AI-assisted surgery. Techniques such as fluorescence-guided surgery, hyperspectral imaging, and augmented reality can be integrated with AI algorithms to provide real-time information on tissue perfusion, tumor margins, and surgical anatomy[39-41]. These innovations have the potential to improve the precision of surgical resection, optimize oncologic outcomes, and preserve postoperative function.

Beyond direct patient care, surgical data science offers opportunities for the objective assessment of surgical performance. ML algorithms can be trained to evaluate technical skill, efficiency, and procedural consistency using video recordings and instrument kinematic data[36,37]. These approaches provide objective, reproducible, and unbiased measures of surgical proficiency that can be used for resident training, competency assessment, and credentialing of practicing surgeons. Ultimately, such applications may contribute to improved surgical education and higher-quality patient care.

Despite significant advances in the development of clinical image analysis algorithms and computer-assisted diagnostic technologies, they have not been widely translated into routine clinical practice[42-44]. Variability in surgical data quality and character represents a significant challenge. For example, images obtained by a high-definition cameras vs those obtained from standard endoscopes, variability in surgical technique, differences in technology use between centers, can all affect performance. Furthermore, additional challenges include applying a model validated on a small retrospective dataset in real time to clinical care, which typically requires a robust IT infrastructure integration with existing electronic health record systems and surgeon training. However, consideration must also be given to the regulatory and ethical issues concerning patient data privacy, AI algorithm transparency and accountability, and medico-legal liability. These challenges can be addressed by technology, technical solutions, and translational research. Table 1 outlines the current limitations and how they could be addressed to realize the clinical benefits of intraoperative AI.

INTERPRETABILITY AND EXPLAINABLE ARTIFICIAL INTELLIGENCE

Although high-performance deep learning-based clinical models can achieve remarkable predictive accuracy, they often function as “black boxes”, limiting clinical trust and integration into practice[45]. This limitation is particularly pronounced in dynamic settings such as surgical oncology, where decisions derived from opaque models are difficult to interpret and therefore challenging to apply. For effective clinical decision support, it is essential that clinicians understand the reasoning underlying the models to adapt their outputs to specific medical contexts and requirements[46]. To ‘open the black-box’ of deep learning techniques, new methods have emerged, known as explainable AI (XAI), which have recently been integrated into risk models[45]. Here, SHapley Additive exPlanations and Local Interpretable Model-Agnostic Explanations are used to decompose the contribution of individual inputs to predictions[47]. This allows clinicians to understand the key variables driving risk estimates for individual patients. In gastric cancer surgery, many factors are considered during the decision-making process, including tumor-related, host-related and perioperative factors. Interpretation of deep learning predictions using XAI methods helps clinicians understand the extent that deep learning model-based decision-making is supported by established biological and clinical knowledge, thereby facilitating clinical validation and acceptance[48].

In addition to post hoc model interpretation techniques that quantify the contribution of individual variables to model predictions and facilitate clinical decision-making, a variety of approaches have been developed to enhance XAI. These methods aim to address important concerns related to model transparency, causality, fairness, and potential sources of bias, thereby improving the trustworthiness and clinical applicability of AI-driven predictive models[49].

Most existing methods for explaining deep learning models are post-hoc in nature. However, many current feature-attribution approaches have limited fidelity and are inherently unstable, which can reduce their reliability across different datasets and model architectures and hinder reproducibility[50]. Furthermore, that lack of interpretability should not be equated with a lack of causality. While XAI methods are valuable for identifying associative relationships and clarifying how specific variables contribute to model predictions, these approaches generally do not establish causal effects. In surgical decision-making, understanding the underlying causal mechanisms that drive outcomes is often more important than identifying correlations alone[51].

A further challenge in the development of AI models is the potential for bias arising from imbalanced or non-representative training datasets. Although XAI techniques can help identify potential sources of bias by visualizing the influence of individual variables on model predictions, they do not eliminate the underlying bias in the model itself.

From a translational perspective, the development of interpretable predictive models is highly desirable. Generalized additive models (GAMs) offer a balance between predictive performance and interpretability, while rule-based systems can also achieve strong predictive accuracy while maintaining transparency. In addition, hybrid approaches that combine high-performing ML models with interpretable components, such as decision trees, may enhance clinical usability[52].

AI has the potential to improve the delivery and outcomes of cancer surgery. However, broader clinical adoption will require models that not only achieve high predictive performance but also provide more robust image interpretation and clinically meaningful explanations. These considerations are summarized in Table 1.

MULTIMODAL DATA INTEGRATION IN ARTIFICIAL INTELLIGENCE-DRIVEN SURGICAL ONCOLOGY

The integration of multiple data modalities within AI-assisted surgical oncology represents one of the most promising frontiers in the field. Contemporary AI- and ML-based prognostic models have expanded beyond traditional clinical and pathological variables to incorporate diverse data sources, including imaging, genomic, and intraoperative information. By leveraging these complementary datasets, such models can capture complex biological and clinical relationships that would be difficult to characterize using any single data modality alone. This approach facilitates the simultaneous consideration of interactions between the tumor and its host, as well as dynamic relationships between the tumor, treatment and clinical outcomes[53,54].

Multimodal learning has attracted increasing interest in recent years. There is also growing interest in incorporating time-dependent information, such as postoperative recovery data (e.g., vital signs) and follow-up imaging. By learning from temporal data acquired during the entire perioperative period in a single task, one can move from static prediction to dynamic risk modeling, where a ML model updates the risk profile of a patient over time and during his/her recovery from surgery, using data acquired during recovery and follow-up imaging[55,56].

While there are many benefits to integrating different data types, there are several challenges to data harmonization, feature selection, and model interpretation that must be overcome when combining several sources of heterogeneous data, with heterogeneity arising from different imaging protocols, sequencing platforms, or clinical data collection. When developing a computer vision-based disease diagnosis system, an important problem to address is how to fuse multimodal features together. Some existing methods directly fuse all the raw multimodal data (multimodal early fusion) into models before learning, while others fuse the output of different models that have been trained on individual modalities (multimodal late fusion)[57]. While hybrid fusion methods have also been explored, there is still no optimal solution for fusing multimodal data in clinical settings.

Despite considerable advances in deep learning, the major impediments to the wider clinical dissemination of neural networks remain a lack of interpretability and excessive computational time. Furthermore, as the dimensionality of the feature space increases, like when incorporating multiple inputs, the risk of overfitting in small training datasets used in surgical oncology becomes significantly greater[15,58]. Figure 2 shows how diverse data sources, including clinical, pathological, radiological, and genomic information, are processed to generate extracted insights such as imaging features, histopathological patterns, and molecular variants. These heterogeneous data are subsequently integrated through artificial intelligence models to support key clinical applications, including diagnosis, treatment selection, prognosis, tumor typing, and postoperative monitoring. This framework highlights the role of multimodal AI in enabling comprehensive and personalized decision-making in gastric cancer management.

Figure 2 Multimodal data integration framework in artificial intelligence-driven surgical oncology. The figure was created with assistance from ChatGPT for visual design and language support. However, the scientific content, interpretation, and overall concept of the figure were developed entirely by the authors. AI: Artificial intelligence.

In addition to developing robust AI systems that can be effectively translated into the clinical environment, there is a need for reliable clinical data pipelines and strong communication between clinicians, data scientists, and engineers. Equally important is the establishment of robust implementation frameworks.

EXTERNAL VALIDATION AND CLINICAL IMPACT OF ARTIFICIAL INTELLIGENCE MODELS

A major limitation to the application of ML to gastric cancer surgery is the lack of adequate external validation. Although existing models have been developed and internally validated using patient data from a single surgical center, independent validation and verification of their performance on patients from different centers using high-quality external clinical sets are needed[12,59]. External generalization to independent multicenter cohorts and prospective assessment is critical to evaluate utility in real world use cases.

While predicting potential cancer progression or recurrence is an important first step, it is not sufficient to ensure that such systems make it into clinical practice[60]. The next wave of research will be to investigate the impact of such computer-aided systems on the clinical decision-making process, as well as patient and system outcomes.

COMPARISON BETWEEN ARTIFICIAL INTELLIGENCE AND CLINICAL DECISION-MAKING

Several studies have compared the performance of ML algorithms with that of experienced clinicians in predicting surgical outcomes[13,61-63]. In many cases, AI models demonstrate performance comparable to or exceeding that of clinicians, particularly when handling complex, high-dimensional datasets with multiple interacting variables. As the volume and complexity of clinical data continues to increase, ML is emerging as a valuable adjunct in surgical decision-making, capable of rapidly processing large datasets and identifying non-linear patterns that may not be readily apparent to human observers.

Despite this promise, important limitations must be acknowledged. Most comparative studies are conducted under controlled conditions using high-quality datasets, whereas real-world clinical decision-making involves uncertainty, incomplete data, and time constraints, along with a range of contextual, ethical, and patient-specific considerations[12]. Importantly, improving predictive accuracy does not equate to replicating clinical decision-making, which requires balancing risks and benefits while incorporating patient preferences and quality-of-life considerations[64].

Furthermore, the evaluation of AI models is typically based on performance metrics such as area under the curve and accuracy. However, high predictive performance does not necessarily translate into clinical utility, particularly if model output does not meaningfully influence management decisions or improve patient outcomes. Therefore, evaluation frameworks should shift from prediction-level metrics toward decision-level impact.

Generalizability represents another critical challenge. ML models are inherently dependent on the data on which they are trained and may perform poorly when applied to patient populations that differ from the original dataset. This limitation is especially relevant in surgical oncology, where patient characteristics, surgical techniques, and institutional practices vary widely. In contrast, clinicians can adapt their decision-making strategies across diverse scenarios, although human judgment is also subject to cognitive biases, interobserver variability, and fatigue-related errors. AI systems offer advantages in terms of reproducibility, consistency, and scalability, which may help mitigate some of these limitations. However, they can also introduce systematic biases embedded within training datasets if not carefully validated and monitored. For this reason, AI should not be viewed as a replacement for clinical expertise, but rather as a complementary decision-support tool that informs human judgment[65].

Future clinical decision-support systems should adopt hybrid models that integrate the strengths of both AI and clinicians. Emerging evidence suggests that human-AI collaboration may achieve higher diagnostic accuracy than either approach alone[66]. Accordingly, future research should focus on optimizing collaborative frameworks and evaluating their real-world impact on clinical decision-making and patient outcomes.

Ultimately, the successful integration of AI into surgical oncology will depend on the development of accurate, interpretable, and clinically relevant systems that align with real-world decision-making processes. Addressing current limitations and bridging existing gaps will be essential to enable the effective translation of AI into routine clinical practice.

ETHICAL AND REGULATORY CONSIDERATIONS

The integration of AI in surgical oncology raises critical ethical and regulatory challenges that must be addressed to ensure safe and equitable clinical implementation. Key concerns include data privacy, algorithmic bias, and transparency, particularly given the increasing reliance on large-scale clinical datasets derived from electronic health records, imaging systems, and genomic databases. These data sources introduce important considerations regarding data ownership, informed consent, and secure data storage, all of which must comply with local and international data protection regulations to safeguard patient confidentiality[67-70].

Regulatory frameworks governing AI-based medical devices are evolving rapidly but remain insufficiently standardized. Most current approval pathways were designed for static systems, whereas many contemporary AI models are adaptive and continuously updated. This raises important questions regarding the frequency of re-validation, mechanisms for monitoring post-deployment model updates, and the establishment of criteria to ensure sustained performance and safety over time[71-73].

Algorithmic bias represents a major concern in surgical oncology, where non-representative training datasets may lead to unequal outcomes across patient populations. Therefore, beyond predictive accuracy, ensuring dataset diversity and implementing continuous bias monitoring are essential. While transparency and explainability are necessary to support clinical trust, they are not sufficient on their own; models must also undergo rigorous clinical validation and be tailored to specific clinical use cases.

In addition, unclear medico-legal responsibility among clinicians, healthcare institutions, and developers continues to hinder widespread adoption. Addressing these challenges requires not only regulatory oversight but also the development of robust data infrastructures, standardized validation frameworks, and targeted clinician training programs. Continuous performance monitoring and governance mechanisms will be essential to ensure reliability in real-world settings.

Ultimately, the deployment of AI in surgical oncology must be guided by principles of fairness, accountability, and transparency, ensuring that technological innovation translates into clinically meaningful and ethically sound improvements in patient care.

CLINICAL IMPLICATIONS AND FUTURE PERSPECTIVES

The study by Lü et al[9] established TNM stage, lymph vascular invasion, age, tumor diameter, albumin, carcinoembryonic antigen, and intraoperative blood loss as prognostic factors. These findings are consistent with previous studies demonstrating that each of these variables is associated with long-term outcomes in patients undergoing surgery for gastric cancer[6,7]. The features incorporated into this model capture distinct aspects of postoperative risk rather than reflecting a single prognostic factor.

The model is well suitaed for real-world application because it is implemented through a simple, web-based platform that is readily accessible to users. By enabling real-time, individualized risk assessment using routinely collected clinical data, the authors have taken an important step toward improving postoperative decision-making in gastric cancer care. Such tools can support the development of risk-adapted surveillance strategies and inform decisions regarding adjuvant therapy. Despite this promise, several barriers may limit the widespread adoption of web-based AI tools in clinical practice. Challenges include integration with electronic health record systems, clinician training programs, establishment of data governance standards, and navigation of regulatory clearance procedures. The successful implementation of prediction models in surgical oncology requires solutions that solve translation-related problems to achieve enduring meaningful results.

Several important challenges remain to be addressed before AI-based prognostic models can be widely adopted in clinical practice. External validation in multicenter and prospective cohorts is essential to establish the generalizability of these findings across diverse population groups and healthcare settings. Future research should also focus on integrating molecular, genomic, and immune biomarkers into predictive models, which may further improve prognostic accuracy while providing deeper insights into the biological mechanisms underlying disease progression and treatment response.

CONCLUSION

In conclusion, the study by Lü et al[9] demonstrates the growing clinical relevance of AI-based prognostic modeling in gastric cancer surgery. By integrating advanced missing data handling, interpretable ML techniques, and a web-accessible prediction interface, the proposed framework represents an important step toward clinically deployable decision-support tools. Nevertheless, the retrospective single-center design and the lack of external validation or molecular biomarkers remain key limitations. Future research incorporating multicenter prospective validation and biologically integrated datasets will be essential to fully realize the potential of interpretable AI in personalized postoperative management after curative gastrectomy.