Exploring the Potential of Machine Learning in Gastric Cancer: Prognostic Biomarkers, Subtyping, and Stratification

Exploring the Potential of Machine Learning in Gastric Cancer: Prognostic Biomarkers, Subtyping, and Stratification | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Exploring the Potential of Machine Learning in Gastric Cancer: Prognostic Biomarkers, Subtyping, and Stratification Haniyeh Rafiepoor, Mohammad Mehdi Banoei, Alireza Ghorbankhanloo, and 4 more This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-5731247/v1 This work is licensed under a CC BY 4.0 License Status: Published Journal Publication published 30 Apr, 2025 Read the published version in BMC Cancer → Version 1 posted 7 You are reading this latest preprint version Abstract Background Advancements in the management of gastric cancer (GC) and innovative therapeutic approaches highlight the significance of the role of biomarkers in GC prognosis. Machine-learning (ML)-based methods can be applied to identify the most important predictors and unravel their interactions to classify patients, which might guide prioritized treatment decisions. Methods A total of 140 patients with histopathological confirmed GC who underwent surgery between 2011 and 2016 were enrolled in the study. The inspired modification of the partial least squares (SIMPLS)-based model was used to identify the most significant predictors and interactions between variables. Predictive partition analysis was employed to establish the decision tree model to prioritize markers for clinical use. ML models have also been developed to predict TNM stage and different subtypes of GC. Latent class analysis (LCA) and principal component analysis (PCA) were carried out to cluster the GC patients and to find a subgroup of survivors who tended to die. Results The findings revealed that the SIMPLS method was able to predict the mortality of GC patients with high predictabilities (Q 2 = 0.45–0.70). The analysis identified MMP-7, P53, Ki67, and vimentin as the top predictors. Correlation analysis revealed different patterns of prognostic markers in the non-survivor and survivor cohorts and different GC subtypes. The main prediction models were verified via other ML-based analyses, with a high area under the curve (AUC) (0.84–0.99), specificity (0.82–0.99) and sensitivity (0.87–0.99). Patients were classified into three clusters of mortality risk, which highlighted the most significant mortality predictors. Partition analysis prioritizes the most significant predictors P53 ≥ 6, COX-2 > 2, vimentin > 2, Ki67 ≥ 13 in mortality of patients (AUC = 0.85–0.90). Conclusion The present study highlights the importance of considering multiple variables and their interactions to predict the prognosis of mortality and stage in GC patients through ML-based techniques. These findings suggest that the incorporation of molecular biomarkers may enhance patient prognosis compared to relying solely on clinical factors. Furthermore, they demonstrate the potential for personalized medicine in GC treatment by identifying high-risk patients for early intervention and optimizing therapeutic strategies. The partition analysis technique offers a practical tool for identifying cutoffs and prioritizing markers for clinical application. Additionally, providing Clinical Decision Support systems with predictive tools can assist clinicians and pathologists in identifying aggressive cases, thereby improving patient outcomes while minimizing unnecessary treatments. Overall, this study contributes to the ongoing efforts to improve patient outcomes by advancing our comprehension of the intricate nature of GC. Gastric cancer Prediction model Machine learning Immunohistochemistry Mortality Figures Figure 1 Figure 2 Figure 3 Figure 4 Figure 5 Introduction Gastric cancer (GC), accounting for 7.7% of total cancer deaths, is the third leading cause of cancer deaths worldwide. According to estimates published by the International Agency for Research on Cancer (IARC), over 750,000 new mortalities of GC worldwide and more than 12,500 new cases of mortality have been reported in Iran. [ 1 ] Despite all the advantages in the field of diagnosis and treatment of GC, the prognosis of patients with GC is still very poor, especially the long-term survival of patients with advanced GC, for whom the median survival is less than 12 months [ 2 ]. Therefore, accurate staging of GCs in patients may improve the management and outcome of these patients [ 3 ]. The TNM staging system provided by the International Union against Cancer/American Joint Committee on Cancer (UICC/AJCC)) has been recognized as the most important staging system [ 4 ]. However, the TNM staging system has limited applications, and middle-stage patients present diverse prognostic outcomes, making it difficult to accurately predict the mortality of patients. [ 5 ]. Advancements in the management of GC and innovative therapeutic approaches offer preoperative treatment options and highlight the importance of identifying high-risk patients and the role of biomarkers in GC management [ 6 , 7 ]. These factors underscore the need for the development of new and accurate prognostic models [ 8 , 9 ]. Using various artificial intelligence (AI) algorithms, nonlinear statistical models can be constructed to predict the survival of GC patients and categorize patients into groups with better distinguishing abilities via unsupervised machine learning (ML) algorithms [ 10 – 13 ]. In the past few years, several models with different capabilities for determining the prognosis of GC have been designed [ 9 , 10 , 13 – 17 ]. Numbers of these models include only common clinical indicators (TNM stage, age, sex, etc.) to determine prognosis [ 14 – 16 ], and few models consider biomarkers in addition to other clinical factors, which increases the accuracy of the models. However, the number of these biomarkers is limited and can be expanded [ 9 , 13 , 17 ]. Additionally, the performed analysis has a limited ability, which can be solved by using AI and ML algorithms. Traditional analysis methods often have limitations in handling complex and multidimensional datasets, identifying nonlinear patterns, processing unstructured data, and performing predictive analysis. AI and ML algorithms can address these limitations and provide more robust and efficient solutions [ 18 ]. The study of multiple biomarkers and their interactions plays a crucial role in understanding the complex relationships among molecular pathways, particularly in diseases such as gastric cancer. Furthermore, the potential interactions among biomarkers are often overlooked in conventional models, impeding the complex relationships among molecular pathways. Notably, the integration of molecular pathology and AI could lead to novel biomarkers that have diagnostic or prognostic value. Several biomarkers, such as human epidermal growth factor receptor 2 (HER-2) [ 19 ], matrix metalloproteinase 7 (MMP-7) [ 20 , 21 ], cyclooxygenase 2 (Cox2) [ 22 , 23 ], vimentin [ 24 , 25 ], tumor protein 53 (TP53) [ 26 ], CD34 [ 27 ], and Ki67 [ 28 ], have shown potential as markers for angiogenesis and cell proliferation in the prognosis of patients with GC. In the present study, our objective was to utilize an AI model to identify potential biomarkers and their combinations for prognostic purposes and variant classifications in gastric cancer. An ML-based statistical method was applied to predict the mortality of GC patients and identify the complex interactions between predictors. The clustering method was used to design and classify the patients into different mortality groups (low-, moderate-, and high-risk groups) and to understand the role of each feature in the groups. To achieve this objective, a combination of seven biomarkers and other clinical and histopathological factors was employed in a cohort of GC patients. Identifying potential predictors is helpful for determining the appropriate prognosis and improving the understanding of the pathophysiological behavior of GC. Methods and Materials Data collection A retrospective cohort of 140 patients diagnosed with GC was included in the study, which was previously described by Razavirad et al. [ 29 ], who traditionally analyzed the pathological impacts of biomarkers. Clinical data and histopathological findings from patients were collected at the Cancer Institute, Imam Khomeini Hospital Complex, Tehran University of Medical Sciences, I.R. of Iran, from April 2011 to January 2016. All patients who underwent neoadjuvant therapy were excluded from the study. The pathologic stage of patients was classified according to the 8th edition of the American Joint Committee on Cancer (AJCC) classification. Patients with stage I–III disease (n = 140) were analyzed, and the follow-up period was 60 months. A total of 21 variables, including patient demographics, histological findings and 7 biomarkers, were collected from pathologically confirmed patients with GC. Formalin-fixed, paraffin-embedded primary tumor specimens were retrieved from the Cancer Institute archives, sectioned into 3 µm slices, and processed for immunohistochemistry (IHC). Tissue sections were deparaffinized, antigen-retrieved in citrate buffer, blocked with bovine serum albumin, and incubated with specific antibodies against HER2, CD34, p53, Ki67, COX-2, MMP7, and vimentin. Immunoreactivity was assessed by two pathologists, with HER2 scored per CAP guidelines, CD34 evaluated for microvessel density, and p53, Ki67, COX-2, MMP7, and vimentin categorized based on staining intensity [ 30 ]. In this study, participants with missing values were excluded to minimize bias and ensure the accuracy of the ML models. Mortality prediction was performed at two time points, 18 and 24 months, as the median and mean survival times for GC patients, which is consistent with previous studies in Iran [ 31 ]. However, the current cohort was followed for 5 years, and only 4% of patients survived. There was a significant group imbalance between survivors and non-survivors from 24 months until 5 years. Because we have a relatively small sample size, group imbalance can lead to the development of biased classifiers, where the predictive performance is skewed toward the majority class, potentially resulting in accurate predictions for the minority class. Continuous and ordinal variables were converted to binary variables via a categorical approach (cutoff point for each variable) using partition analysis (JMP pro, SAS). This study was approved by the ethics committee of Imam Khomeini Hospital Complex- Tehran University of Medical Sciences, Tehran, Iran (No. IR.TUMS.IKHC.REC.1400.001). Statistical analysis The statistically inspired modification of partial least squares (SIMPLS), an algorithm of PLS (a linear ML method) [ 32 , 33 ] analysis, was applied to create a prediction model of 18- and 24-month mortality via patient demographics, histopathological findings and biomarker variables. SIMPLS-based prediction models were obtained through the training and validation sets. The most differentiating variables with variable importance in the projection (VIP) > 1.0 were selected for the building prediction model. The validation set was randomly selected on the basis of 30% of the 140 patients with GC, which was also considered internal validation. Q 2 and R 2 Y were counted as the goodness for predictability and the goodness of variability, respectively, to evaluate the performance of the SIMPLS-based prediction model. The best prediction models were obtained using the most differentiating variables when the Q 2 reached the highest value before decreasing and with the highest R 2 Y. Q 2 and R 2 Y were computed and verified through the training and validation sets, respectively. The performances of the models are evaluated via leave-one-out cross validation (LOOCV) and accuracy. To categorize the continuous and ordinal values, partition analysis was performed via a decision tree to find the best cutoff point of variables on the basis of the relationship between the outcome and predictors. Latent class analysis (LCA) was performed to identify low- and high-risk patients with GC through the clustering of cohorts using the most differentiating variables. Principal component analysis (PCA) was used to identify the different clusters obtained via LCA. PCA was also applied to find the trend and outliers using all the variables. Cox regression and Kaplan–Meier survival analyses were performed on the AJCC stage groups (stages I, II, and III in the favorable category) among the test data and our model to compare the distinguishing ability between the two methods. Model screening was conducted to check the final prediction model by providing a summary table via other ML methods, such as XGBoost, support vector machine (SVM), boosted tree (BT), bootstrap forest (BF), K-nearest neighbor (KNN), generalized regression lasso (GRL), boosting neural network (BNN), fit stepwise (FS), and naïve Bayes (NB) methods. Model screening also helps in finding efficient workflows as well as comparing and exploring datasets for the best predictive model. We further applied predictive partition analysis (PPA) or a decision tree, which is ML method [ 34 ] that could be used to classify or predict data responses. The ML aspect of the decision tree comes into determining when and where to divide the data between branches. Data partitioning was performed by dividing a dataset into subsets, including training and validation groups. Results Patient characteristics A total of 140 patients with confirmed GC were enrolled in the study, and 62 (44%) and 99 (70%) patients died within 18 and 24 months, respectively. Table 1 shows the demographic characteristics and histopathological and biochemical markers of survivors and non-survivors GC patients at 18 and 24 months after disease onset. The patient information is available in supplementary table E1. Predicting Mortality in Gastric Cancer Patients via Clinical and Paraclinical Data An overview of the present study is shown in Figure 1. The ML-based approach showed that demographics, histopathological and biochemical markers can be used for predicting the mortality outcomes of patients with GC. SIMPLS analysis was carried out via most differentiating variables (VIPs) [35] to establish the prediction model. The prediction model was developed on 95 patients in the training set and 45 patients in the validation set. Two factor-based SIMPLS models had high predictabilities (Q 2 = 0.45 and Q 2 = 0.70) for predicting mortality at 18 and 24 months, respectively, and included a total of 10 variables that contributed to the prediction models. SIMPLS-based scatter plots demonstrated very good discrimination between survivors and non-survivors for both 18- and 24-month mortality prediction studies (Figure 2. A & C). Table 2 shows that MMP-7, P53, and Ki67 were the most important variables for 18-month mortality. For 24-month P53, Ki67 and vimentin were considered the top 3 predictors associated with 24-month mortality in patients with GC. Furthermore, the coefficient plots revealed that pathological tumor scoring and the presence of regional lymph nodes, vimentin, HER-2, COX, MMP-7, Ki67, and P53 were positively correlated with 18- and 24-month mortality. Low CD34 and high CD34 levels were correlated with mortality at 18- and 24-month mortality, respectively, indicating that the role of CD34 may change in mortality over time (Figure 2. B & D). PPA verified that the abovementioned most differentiating variables are strong predictors for mortality at 18 and 24 months for the training and validation sets in comparison with TNM staging (Table 3). PCA-based correlation analysis showed that the abovementioned differentiating variables were highly correlated with each other for 18- and 24-month mortality prediction (Figure 3). Age group was less correlated with most of the variables. Additionally, tumor size was positively correlated with regional lymph node variables in both 18- and 24-month mortality studies. PCA correlations among the most important variables are available in supplementary table E2. Model screening revealed high AUCs (> 0.80), high specificities (> 90%), and good sensitivities (>70%) when the most differentiating predictors were used in other ML methods, such as SVM, KNN, and GRL (Table E3). Identification of high - and low-risk patients with gastric cancer Further investigations using LCA showed that patients with GC can be clustered to identify high-risk patients based on the clinical and biomarker data. LCA-based clustering revealed three main clusters among survivors and non-survivors. LCA-based clustering revealed that cluster 2 and cluster 3 had 44% and 70% mortality rates, respectively (Figure 4). Compared with Clusters 2 and 3, Cluster 1 had the lowest rate of mortality (0%). All 3 clusters were well depicted through a PCA plot that can verify the clustering via two unsupervised methods. Table 4 shows that although variables had different contributions (conditional probabilities) to each cluster, several variables markedly impact clustering. Hence, CD34 ≥ 30, P53 ≥ 6, Ki67 ≥ 13, and vimentin >2 were highly correlated with cluster 3 and the highest rate of mortality. On the other hand, the biomarkers MMP-7 ≤ 2, HER-2 ≤ 1, pathological tumor 1, regional lymph node > N1, COX-2 >2 and MMP-7 >2 showed similar probabilities for clusters 2 and 3. Cox regression analysis and Kaplan‒Meier survival curves revealed that the survival curve of each subgroup had good precision for 18- and 24-month mortality ( p < 0.05). For both the clinical TNM staging and LCA clustering methods, the curves showed acceptable discrimination ( p 2 and vimentin >2 are associated with the prediction of mortality, and patients with P53 <6 and vimentin ≤ 2 are associated with survival (Supplementary Figure E2). Also, 24-month-based partition analysis showed that P53 ≥ 6 and Ki67 ≥ 13 in patients with GC were associated mainly with mortality. On the other hand, patients with P53 <6 and regional lymph nodes ≤ N1 were associated with survival outcomes (Supplementary Figure E3). P redicting TNM stages of gastric cancer A SIMPLS-based prediction model revealed that vimentin, P53 and HER-2 can predict the TNM stage of GC patients as an outcome with good predictability (Q2 = 0.45) (Figure 5C). These biomarkers were the most differentiating among demographics and histopathological and biochemical markers when used in dichotomized values. The predictive decision tree demonstrated cut-off points of vimentin (≤ 2 or > 2), P53 ( 1) for the diagnosis of TNM-based staging, which was interestingly the same for the prediction of mortality outcome (Supplementary Figure E4). Model screening showed high AUCs (> 0.80), high specificities (> 90%), and good sensitivities (>70%) when the most differentiating predictors were used in other ML methods, such as SVM, KNN, and GRL (Supplementary Table E4). Decision tree analysis also revealed that vimentin is the most differentiating biomarker, whereas vimentin 3 were mostly associated with TNM stages ≤ IIb and > IIb, respectively (Supplementary Figure E4). Prediction of Different Types of Gastric Cancer: Tumor Location and Histology We applied the SIMPLS method to predict the tumor location (cardiac vs. non-cardiac) and tumor histology (intestine vs. non-intestine, diffuse vs. non-diffuse) as well as the tumor type (adenocarcinoma vs. non-adenocarcinoma) using clinical and biomarker variables. The models for differentiating adenocarcinoma from non-adenocarcinoma, and the intestine from the non-intestine were acceptably predictive (Q 2 = 0.33 and Q 2 = 0.34, respectively). Tumor histology, histology grade, tumor size, and age mostly contributed to predicting adenocarcinoma. Adenocarcinoma, pathological tumor, histology grade, HER-2 status, and age contributed to the ability to predict the tumor histology of the intestine from that of the non-intestine. The coefficient plot indicates how variables are correlated with the GC intestine and non-intestine subtypes (Figure 5A-B). Multivariate correlation analysis showed higher correlation between biomarkers in the cardia and intestine subtypes than between biomarkers in the noncardia and non-intestine subtypes. Age was negatively correlated with other variables among non-cardia subtypes compared with cardia. This phenomenon was also visible in the intestine and non-intestine subtypes (Supplementary Figure E5). Discussion Despite all the progress that has been made in the field of biomarker identification and model development in the field of oncology, there are still limitations in classifying patients in terms of prognosis and identifying the roles of various indicators and their relationships. In this study, the findings indicated that the ML-based SIMPLS model can accurately predict the mortality of GC patients via demographic and histopathological data, along with identifying the most important predictors and interactions between variables at two time points. The mean survival time for the patients was 18 months. The study focused on comparing the mortality rates between 18 and 24 months at 6-month intervals. In our study, 9 variables, including 2 clinical and 7 histopathological markers, were identified as potential prognostic predictors (Table 2 ). Mortality predictor variables were weighed and ordered on the basis of their importance in the prediction model. Hence, for 18-month mortality, MMP-7 > 2, P53 ≥ 6%, and Ki67 > 13 and for 24-month mortality, P53 ≥ 6%, Ki67 > 13, and vimentin > 2 were regarded as the top three predictors with the greatest impact on the model in the presence of other clinical risk factors, such as vascular invasion. This observation highlights the capacity of molecular biomarkers to provide a more accurate prognosis than do clinical factors. Nonetheless, the model demonstrated lower efficacy in the absence of additional predictors, as indicated in Table 1 . Few numbers of studies have been conducted on the development of AI-based models in the field of GC prognosis. Some studies have utilized deep learning algorithms for image analysis [ 36 – 38 ], whereas others have used numerical algorithms [ 9 , 10 , 12 – 17 , 39 – 41 ], most of them have used SVM and ANN algorithms for model development and reported accuracies ranging from 0.79–0.94. Some of the studies included demographic (age, sex) and clinicopathological (tumor size, pathological grade, TNM staging) features and two biomarkers (CEA, CA199) as predictors. However, in the study of Jiang et al. [ 40 ], in addition to 3 baseline features (sex, CEA, and lymph node metastasis), 8 IHC markers (CD3invasive margin, CD3center of tumor, CD8IM, CD45ROCT, CD57IM, CD66bIM, CD68CT and CD34) were also considered. The main objective of most studies was to design a model using ML approaches to predict patient mortality with higher accuracy, which is traditionally predicted by TNM staging. TNM staging has been recognized as the most crucial staging system for stratifying GC patients into different risk groups [ 4 ]. However, middle-stage patients have a variety of prognostic outcomes, and a more accurate classification of these patients is required [ 5 ]. Importantly, the TNM staging system is applicable only after surgery due to the requirement of pathological analysis of tumor specimens and examination of lymph nodes. Nevertheless, some studies have suggested new patient classification methods to improve the TNM staging system [ 4 , 42 , 43 ]. In the study of Razavirad et al. [ 29 ], the data were analyzed using traditional statistical analysis methods. The present study employed unsupervised ML techniques, PCA and LCA, to cluster patients. This approach differs from previous studies, such as those of Que SJ et al. [ 9 ] and Oh et al. [ 13 ], who utilized supervised ML algorithms to categorize patients into mortality groups based on predicted survival probability. In the unsupervised learning approach, unique classes can be identified, which might offer a superior classification method in comparison to previous approaches [ 44 ]. The importance of each feature in each cluster was also revealed through this unsupervised approach. By considering the incorporation of all the parameters in model development and a thorough discussion of the statistical process, this study has shed light on the importance of individual factors and their interplay with other variables in clinical and statistical aspects that were not addressed in previous studies. Given the complexity of AI analysis, it is beneficial to consider the associations between features and the formation of distinct clusters to enhance medical comprehensibility. Furthermore, identifying the interaction of parameters can offer insights into GC molecular pathways [ 45 ]. The correlations and interactions between different variables are presented in Fig. 3 , revealing significant links among biomarkers such as vimentin/CD34, Ki67/CD34, and p53/Ki67 in non-survival cases. In the survival groups, notable correlations were observed between vimentin/COX2 and MMP-7/p53, suggesting the possibility that similar molecular pathways contribute to these relationships. Additionally, alterations in correlations were detected during different periods for certain biomarkers, including HER-2/vimentin, HER-2/CD34, vimentin/Ki67, and Ki67/CD34. Several studies have shown significant relationships between these markers, such as the coexistence of positive HER2 status and vimentin expression, indicating more aggressive proliferation of carcinoma cells [ 46 ]. In the study of Ludovini et al. [ 47 ], the significant relationship between HER2 overexpression and CD34 has been identified as an important indicator reflecting microvessel density (MVD). A combination of p53/MMP-7 expression was found to be a prognostic indicator for Stage II/III GC. [ 48 ] The relationship between p53 expression and Ki-67 was not statistically significant, but the combination of p53 and COX-2 status was associated with significantly higher mean Ki-67 values in p53+/COX-2 + tumors than in p53-/COX-2- tumors [ 49 ]. Elevated cytoplasmic expression of TSPAN1 was positively correlated with increased nuclear expression of Ki-67, accompanied by upregulation of MVD, as indicated by the stromal level of CD34 [ 50 ]. Vimentin immunolabeling in the tumor-associated stroma was positively correlated with Ki-67, and COX-2 was found to promote the production of MMP-2, MMP-9, VEGF, and vimentin while inhibiting E-cadherin production [ 51 ]. The relationship between Ki67 and p53 and the histological grade of GC indicates their importance in cell differentiation [ 52 , 53 ], whereas the correlation between p53 and CD34 biomarkers and perineural invasion indicates the role of biomarkers in cancer invasion pathways [ 54 – 57 ]. It has also been shown that Ki67, HER2, vimentin, MMP-7, p53 and CD34 are correlated with gastric cancer mortality, supporting previous works [ 25 , 58 – 62 ]. In contrast to previous studies, the patient prognosis in the current study was assessed at 2 time points, 18 and 24 months. The order of most of the parameters was similar; however, some features, such as MMP-7, differed in importance, which could indicate changes in the pathophysiological role of MMP-7 for 6 months. Differences in the pattern correlations between factors related to 18- and 24-month mortality are also notable for the variability in tumor behavior during this 6-month interval, which requires further investigation to obtain a better understanding of tumor behavior. The role of MMP-7 and other MMPs in tumor growth, invasion and spread can cause this effect [ 63 , 64 ]. In the clinical setting, employing all biomarkers and features to predict patient mortality may be impractical. Consequently, prioritizing the significance of individual biomarkers to select a subset for prognostic purposes can offer a practical approach for accurate prediction. This approach facilitates the early identification of aggressive GC cases and may guide the implementation of different treatment strategies, such as more aggressive chemotherapy or immunotherapy. Also, by classifying patients into distinct risk categories, the study supports a move toward personalized medicine in the treatment of GC. Patients identified within high-risk clusters may benefit from closer monitoring and early intervention, while those in low-risk categories could potentially avoid unnecessary aggressive treatments. Partition analysis revealed that specific biomarker thresholds, p53, COX-2, and pathological tumors for 18-month mortality, and p53, CD34, and regional lymph nodes for 24-month mortality, are indicative of accurate mortality prediction. The results of the partition analysis agreed with those of the SIMPLS prediction model, indicating the importance of the P53 biomarker in the prediction of mortality at different times. Decision tree analysis revealed some differences in the importance of some biomarkers in the prediction of mortality with respect to the VIP scores of biomarkers, which may be related to the difference in the algorithms used for the two methods. Nonetheless, we believe that SIMPLS can be the primary method for identifying the most differentiating variables, and a decision tree could be a secondary method with more applications in clinical settings. Clinical Decision Support (CDS) systems have previously played a significant role in health care informatics and could be further enhanced through the incorporation of decision trees and partition analysis. These methodologies provide clinically practical cutoffs for biomarker expression, making AI-driven prognostic models more interpretable for healthcare providers. Such models can be integrated into clinical decision-support systems, assisting in staging, treatment selection, and the prediction of survival outcomes. Furthermore, these predictive models can be embedded in pathological workflows to assist pathologists in identifying high-risk GC patients. A similar analysis was performed, in which the TNM system was used to identify the most important factors depicted in supplementary figure E3, and interestingly, patients were classified into advanced and early groups according to the conventional definition of advanced and early GC [ 65 ]. While TNM staging relies on tumor size (T), lymph node involvement (N), and metastasis (M), our model integrates multiple biomarkers (P53, Ki67, vimentin, MMP-7, COX-2, HER-2, CD34) alongside clinical factors, achieving higher predictive accuracy (AUC: 0.84–0.99) and better risk stratification. Unlike TNM staging, which is only applicable post-surgery, our model can provide preoperative prognostic information, guiding treatment decisions earlier. Additionally, decision tree analysis identified biomarker cutoff values (e.g., P53 ≥ 6, Ki67 ≥ 13) that can be used in clinical settings. However, while TNM staging is globally accepted and easy to interpret, our model requires external validation and AI explainability techniques before it can be more broadly adopted in clinical settings. Integrating ML-based approaches into clinical workflows could significantly enhance patient prognosis and personalized treatment strategies. Previous investigations discovered that HER-2 status was significantly associated with various clinical parameters, such as tumor depth or pathological tumor (pT) status, recurrence, distant metastasis, and positive lymph node (pN) status, which was supported by prior research [ 66 ]. Moderate to poorly differentiated carcinoma frequently shows P53-positive staining, which is often accompanied by lympho-vascular invasion. Moreover, an analysis based on the pT and pN stages revealed a significant rise in p53-positive cases [ 67 ]. The expression of vimentin was notably elevated in patients with advanced stages of GC, particularly those exhibiting poorly differentiated types of gastric carcinoma [ 46 ]. These findings suggest a potential role for these biomarkers in predicting clinical outcomes and guiding therapeutic decisions for patients with GC. Given the diverse environmental risk factors associated with GC subtypes and the recent development of molecular classifications such as The Cancer Genome Atlas (TCGA) [ 68 ] and Asian Cancer Research Group (ACRG) [ 69 ], it is imperative to comprehend the role of biomarkers in each anatomical (gastroesophageal junction/cardia and non-cardia) [ 70 ] and histological (Laurén classification) [ 71 ] subtype to gain insights into the pathophysiology of GC. This work offers an advantage over previous studies by investigating the differences in biomarker expression between GC subtypes. In intestinal and non-intestinal GC, variation in the expression of HER-2 has been revealed, as has greater expression in the diffuse subtype, which diverges from earlier findings [ 72 ]. In addition, as presented in supplementary figure E5, gastroesophageal junction/cardia GC shows different correlation patterns than non-cardia GC does, suggesting that these seven biomarkers contribute to a molecular pathway similar to that of non-cardia cancer. The present study is subject to certain constraints. First, the sample size is relatively small compared with that of prior investigations. Nevertheless, the SIMPLS model is considered to be a suitable algorithm for cohorts of limited size [ 33 , 73 ]. On the other hand, recent research on various ML techniques has indicated that the effectiveness of ML-based methods is linked to the quality of the data, even with sample size as small as 40–50 [ 74 ]. Additionally, the major limitation of the current study is the lack of an external validation set, which makes the generalizability and robustness of our findings uncertain, as the results have only been assessed within the confines of the current dataset. Additionally, since all patients were from a single institution, potential selection bias may have occurred. Incorporating a large independent validation cohort would have strengthened the study by allowing us to confirm the reproducibility and applicability of our conclusions across different patient populations, thereby enhancing the overall reliability of our findings. The prediction model was developed by dividing the data into training and validation sets and internally validate using the LOOCV method. Further studies are required to repeat the process of data partitioning several times to obtain an average performance. Finally, owing to the low number of patients, it was not possible to stratify them into more than three categories for better patient classification. Despite the 5-year follow-up period, the survival rate of patients at the end of 60 months was only 4%, leading to a significant group imbalance between survivors and non-survivors from 24 months until 5 years. Since we have a relatively small sample size, group imbalance may give rise to the development of biased classifiers, where the predictive performance is skewed toward the majority class, potentially resulting in accurate predictions for the minority class [ 75 ]; therefore, two specific time points, 18 and 24 months as the median and mean survival times, were chosen for conducting a mortality analysis. In conclusion, the present study provides valuable insights into the use of ML-based techniques for prognostic biomarker identification in GC patients. Furthermore, the study classified patients into three categories and identified the major factors associated with mortality and staging in each group. These results indicate that molecular biomarkers could offer a more accurate prognosis compared to relying solely on clinical factors. Specifically, MMP-7, P53, Ki67, and vimentin were identified as the top predictors. The analysis of the correlation between various factors and their interactions revealed that the biomarker relationship pattern could differ among different anatomical and histopathological subtypes of GC, highlighting the possibility of distinct molecular pathways in GC pathogenesis. The partition analysis technique employed in this study offers a practical tool for identifying cutoffs and prioritizing markers to simplify models for clinical use. However, further research with large independent validation cohort is necessary to validate these findings and explore the potential of clinical and histopathological markers for enhancing prognostic accuracy in patients with GC. Future research should focus on validating the identified biomarkers in larger, independent cohorts to assess their reliability and clinical utility. Given that some of these biomarkers are associated with early tumor progression, their validation could contribute to the development of early screening programs, enabling the detection of GC at an earlier, more treatable stage. Additionally, the potential application of these biomarkers in blood-based tests should be explored for non-invasive monitoring of GC progression and recurrence through liquid biopsy approaches. Such advancements could enhance surveillance strategies and improve patient outcomes. Further investigation into the molecular pathways linked to these biomarkers may provide insights into GC pathogenesis and facilitate the development of targeted therapies. Moreover, conducting subgroup analyses based on other factors than histological type, such as age, sex in larger cohorts would enhance disease classification, improve understanding of GC pathophysiology, and refine risk stratification models for personalized treatment approaches. Overall, this study contributes to the ongoing efforts to improve patient outcomes by advancing our comprehension of the intricate nature of cancer. Declarations Acknowledgments We would like to express our sincere gratitude to Dr. Kazem Zendehdel for his invaluable guidance and support throughout the preparation of this study. His expertise in the field of cancer epidemiology has been instrumental in shaping the direction of this work. Availability of data and materials The data is available upon request. Funding A part of the analysis costs of this study was provided by a grant from the Cancer Biology Research Center, the Cancer Research Center, Tehran University of Medical Sciences (Grant No. 52012). Competing of interest The authors declare that they have no competing interests. Ethics approval and consent to participate This study was approved by the ethics committee of Imam Khomeini Hospital Complex- Tehran University of Medical Sciences, Tehran, Iran (No. IR.TUMS.IKHC.REC.1400.001). Written informed consent obtained from all the participants in the manuscript. All experiments were performed in accordance with the Guidelines of Iran National Committee for Ethics in Biomedical Research. Consent for publication Not applicable Author contributions H.R. contributed to the study concept, data cleaning, statistical analysis interpretation and drafting of the manuscript. M.M.B. contributed to the statistical analysis interpretation, and drafting of the manuscript. A.G. contributed to the interpretation and drafting of the manuscript. A.M. contributed to the study design, data collection, patients follow up and preparation. A.R. contributed to the study concept, collection, patients follow up and data cleaning. S.S contributed to the data collection A.S.Z. contributed to statistical analysis and manuscript revision. S.A. contributed to interpretation and drafting of the manuscript, and overall supervision of the study. All authors read and approved the final manuscript. References Bray F, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394–424. Zhang XY, Zhang PY. Gastric cancer: somatic genetics as a guide to therapy. J Med Genet. 2017;54(5):305–12. Janczewski LM, et al. The inaccuracies of gastric adenocarcinoma clinical staging and its predictive factors. J Surg Oncol. 2023;127(7):1116–24. Zhang J, et al. Evaluation of the seventh AJCC TNM staging system for gastric cancer: a meta-analysis of cohort studies. Tumour Biol. 2014;35(9):8525–32. Liu L, et al. The prognostic efficacy of the 8th edition UICC TNM classifications for gastric cancer in Chinese patients: A study based on follow-up system of nursing department. Med (Baltim). 2018;97(36):e12284. Cai H, et al. Advances in molecular biomarkers research and clinical application progress for gastric cancer immunotherapy. Biomark Res. 2022;10(1):67. Chen Y, et al. Predicting response to immunotherapy in gastric cancer via multi-dimensional analyses of the tumour immune microenvironment. Nat Commun. 2022;13(1):4851. Abbas M, et al. Current and future biomarkers in gastric cancer. Biomed Pharmacother. 2018;103:1688–700. Que SJ, et al. Application of preoperative artificial neural network based on blood biomarkers and clinicopathological parameters for predicting long-term survival of patients with gastric cancer. World J Gastroenterol. 2019;25(43):6451–64. Korhani Kangi A, Bahrampour A. Predicting the Survival of Gastric Cancer Patients Using Artificial and Bayesian Neural Networks. Asian Pac J Cancer Prev. 2018;19(2):487–90. Lee J, et al. Deep Learning-Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer. JCO Clin Cancer Inf. 2018;2:1–14. Nilsaz-Dezfouli H, et al. Improving Gastric Cancer Outcome Prediction Using Single Time-Point Artificial Neural Network Models. Cancer Inf. 2017;16:1176935116686062. Oh SE, et al. Prediction of Overall Survival and Novel Classification of Patients with Gastric Cancer Using the Survival Recurrent Network. Ann Surg Oncol. 2018;25(5):1153–9. Amiri Z, et al. Assessing the effect of quantitative and qualitative predictors on gastric cancer individuals survival using hierarchical artificial neural network models. Iran Red Crescent Med J. 2013;15(1):42–8. Biglarian A, et al. Application of artificial neural network in predicting the survival rate of gastric cancer patients. Iran J Public Health. 2011;40(2):80–6. Biglarian A, et al. Determining of prognostic factors in gastric cancer patients using artificial neural networks. Asian Pac J Cancer Prev. 2010;11(2):533–6. Zhu L, et al. Comparison between artificial neural network and Cox regression model in predicting the survival rate of gastric cancer patients. Biomed Rep. 2013;1(5):757–60. Zhang L, et al. From machine learning to deep learning: progress in machine intelligence for rational drug discovery. Drug Discov Today. 2017;22(11):1680–5. Jørgensen JT, Hersom M. HER2 as a Prognostic Marker in Gastric Cancer - A Systematic Analysis of Data from the Literature. J Cancer. 2012;3:137–44. Ii M, et al. Role of matrix metalloproteinase-7 (matrilysin) in human cancer invasion, apoptosis, growth, and angiogenesis. Exp Biol Med (Maywood). 2006;231(1):20–7. Koskensalo S, et al. MMP-7 overexpression is an independent prognostic marker in gastric cancer. Tumour Biol. 2010;31(3):149–55. Lim HY, et al. Increased expression of cyclooxygenase-2 protein in human gastric carcinoma. Clin Cancer Res. 2000;6(2):519–25. Ristimäki A, et al. Expression of cyclooxygenase-2 in human gastric carcinoma. Cancer Res. 1997;57(7):1276–80. Fuyuhiro Y, et al. Clinical significance of vimentin-positive gastric cancer cells. Anticancer Res. 2010;30(12):5239–43. Otsuki S, et al. Vimentin expression is associated with decreased survival in gastric cancer. Oncol Rep. 2011;25(5):1235–42. Wei K, et al. The prognostic significance of p53 expression in gastric cancer: a meta-analysis. J Cancer Res Clin Oncol. 2015;141(4):735–48. Tenderenda M, et al. Expression of CD34 in gastric cancer and its correlation with histology, stage, proliferation activity, p53 expression and apoptotic index. Pathol Oncol Res. 2001;7(2):129–34. Xiong DD, et al. Ki-67/MKI67 as a Predictive Biomarker for Clinical Outcome in Gastric Cancer Patients: an Updated Meta-analysis and Systematic Review involving 53 Studies and 7078 Patients. J Cancer. 2019;10(22):5339–54. Razavirad A. Prognostic significance of HER2, p53, Ki67, MVD- CD34, Cox2, MMP7, and vimentin in gastric cancer. Basic Clin Cancer Res. 2023;15(2):67–80. Park YS et al. A standardized pathology report for gastric cancer: 2nd edition. J Pathol Transl Med, 2023. 57(1): pp. 1–27. Farahmandi F, et al. Survival of gastric cancer patients in Iran: a systematic review and meta-analysis. Gastroenterol Hepatol Bed Bench. 2023;16(3):245–58. Boulesteix AL, Strimmer K. Partial least squares: a versatile tool for the analysis of high-dimensional genomic data. Brief Bioinform. 2007;8(1):32–44. de Jong S. SIMPLS: An alternative approach to partial least squares regression. Chemometr Intell Lab Syst. 1993;18(3):251–63. Charbuty B, Abdulazeez A. Classification Based on Decision Tree Algorithm for Machine Learning. J Appl Sci Technol Trends. 2021;2(01):20–8. Eriksson L, et al. Multi- and Megavariate Data Analysis. Part I Basic Principles and Applications. Second revised and enlarged edition. Ume Sweden: MKS Umetrics AB; 2006. pp. 1–103. Jiang Y, et al. Predicting peritoneal recurrence and disease-free survival from CT images in gastric cancer with multitask deep learning: a retrospective study. Lancet Digit Health. 2022;4(5):e340–50. Jin P, et al. Artificial intelligence in gastric cancer: a systematic review. J Cancer Res Clin Oncol. 2020;146(9):2339–50. Wei T, et al. Survival prediction of stomach cancer using expression data and deep learning models with histopathological images. Cancer Sci. 2023;114(2):690–701. Ishii H, et al. Classification of gastric cancer subtypes using ICA, MLR and Bayesian network. Stud Health Technol Inf. 2013;192:1014. Jiang Y, et al. Immunomarker Support Vector Machine Classifier for Prediction of Gastric Cancer Survival and Adjuvant Chemotherapeutic Benefit. Clin Cancer Res. 2018;24(22):5574–84. Zhou C-M, et al. Predicting postoperative gastric cancer prognosis based on inflammatory factors and machine learning technology. BMC Med Inf Decis Mak. 2023;23(1):53. Chen ZD, et al. Recent Advances in the Diagnosis, Staging, Treatment, and Prognosis of Advanced Gastric Cancer: A Literature Review. Front Med (Lausanne). 2021;8:744839. Lin JX et al. Multicenter Validation Study of the American Joint Commission on Cancer (8th Edition) for Gastric Cancer: Proposal for a Simplified and Improved TNM Staging System. J Cancer, 2020. 11(12): pp. 3483–3491. Liu X. Supervised classification and unsupervised classification. ATS; 2005. Moore JH, Raghavachari N. Artificial Intelligence Based Approaches to Identify Molecular Determinants of Exceptional Health and Life Span-An Interdisciplinary Workshop at the National Institute on Aging. Front Artif Intell. 2019;2:12. Zhao W, et al. Clinical significance of vimentin expression and Her-2 status in patients with gastric carcinoma. Clin Transl Sci. 2013;6(3):184–90. Ludovini V, et al. Evaluation of the prognostic role of vascular endothelial growth factor and microvessel density in stages I and II breast cancer patients. Breast Cancer Res Treat. 2003;81(2):159–68. Sawada T, et al. New molecular staging with G-factor supplements TNM classification in gastric cancer: a multicenter collaborative research by the Japan Society for Gastroenterological Carcinogenesis G-Project committee. Gastric Cancer. 2015;18(1):119–28. Joo YE, et al. Expression of cyclooxygenase-2, p53 and Ki-67 in gastric cancer. J Korean Med Sci. 2006;21(5):871–6. Chen L, et al. Clinicopathological significance of overexpression of TSPAN1, Ki67 and CD34 in gastric carcinoma. Tumori. 2008;94(4):531–8. Zhang X, et al. COX–2 promotes epithelial–mesenchymal transition and migration in osteosarcoma MG–63 cells via PI3K/AKT/NF–κB signaling. Mol Med Rep. 2019;20(4):3811–9. Lazăr D, et al. Ki-67 expression in gastric cancer. Results from a prospective study with long-term follow-up. Rom J Morphol Embryol. 2010;51(4):655–61. Levine AJ, et al. The Role of the p53 Protein in Stem-Cell Biology and Epigenetic Regulation. Volume 6. Cold Spring Harb Perspect Med; 2016. 9. Muller PA, Vousden KH, Norman JC. p53 and its mutants in tumor cell migration and invasion. J Cell Biol, 2011. 192(2): pp. 209 – 18. Yang J, et al. Associations of CD34, Ki67, layer of invasion and clinical pathological characteristics, prognosis outcomes in gastrointestinal stromal tumors—a retrospective cohort study. Translational Cancer Res. 2022;11(8):2866–75. Blanchet A et al. Isoforms of the p53 Family and Gastric Cancer: A Ménage à Trois for an Unfinished Affair. Cancers (Basel), 2021. 13(4). Saito H, et al. Neoangiogenesis and relationship to nuclear p53 accumulation and vascular endothelial growth factor expression in advanced gastric carcinoma. Oncology. 1999;57(2):164–72. Habibi A, et al. Correlation between CD34 Marker and Clinico-Pathologic Characteristics of Gastric Cancer. Int J Cancer Res. 2015;11:136–42. Ko GH, et al. Prognostic impact of Ki-67 in patients with gastric cancer-the importance of depth of invasion and histologic differentiation. Med (Baltim). 2017;96(25):e7181. Soleyman-Jahi S, et al. Prognostic significance of matrix metalloproteinase-7 in gastric cancer survival: a meta-analysis. PLoS ONE. 2014;10(4):e0122316. Uprak TK, et al. HER-2 incidence in gastric cancer, its association with prognosis and clinicopathological parameters. Ulus Cerrahi Derg. 2015;31(4):207–13. Yildirim M, et al. Prognostic significance of p53 in gastric cancer: a meta- analysis. Asian Pac J Cancer Prev. 2015;16(1):327–32. Murray GI, et al. Matrix metalloproteinases and their inhibitors in gastric cancer. Gut. 1998;43(6):791–7. Polistena A, et al. MMP7 expression in colorectal tumours of different stages. Vivo. 2014;28(1):105–10. Hu B, et al. Gastric cancer: Classification, histology and application of molecular pathology. J Gastrointest Oncol. 2012;3(3):251–61. Ognjenovic L, et al. HER2 Positive Gastric Carcinomas and Their Clinico-Pathological Characteristics. Open Access Maced J Med Sci. 2018;6(7):1187–92. Lazăr D, et al. The immunohistochemical expression of the p53-protein in gastric carcinomas. Correlation with clinicopathological factors and survival of patients. Rom J Morphol Embryol. 2010;51(2):249–57. Comprehensive molecular characterization of gastric adenocarcinoma. Nature, 2014. 513(7517): pp. 202–9. Cristescu R, et al. Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes. Nat Med. 2015;21(5):449–56. Siewert JR, et al. [Cardia cancer: attempt at a therapeutically relevant classification]. Chirurg. 1987;58(1):25–32. Lauren P, THE TWO HISTOLOGICAL MAIN TYPES. OF GASTRIC CARCINOMA: DIFFUSE AND SO-CALLED INTESTINAL-TYPE CARCINOMA. AN ATTEMPT AT A HISTO-CLINICAL CLASSIFICATION. Acta Pathol Microbiol Scand. 1965;64:31–49. Wang HB, Liao XF, Zhang J. Clinicopathological factors associated with HER2-positive gastric cancer: A meta-analysis. Med (Baltim). 2017;96(44):e8437. Banoei MM, et al. Unraveling complex relationships between COVID-19 risk factors using machine learning based models for predicting mortality of hospitalized patients and identification of high-risk group: a large retrospective study. Front Med (Lausanne). 2023;10:1170331. Rajput D, Wang W-J, Chen C-C. Evaluation of a decided sample size in machine learning applications. BMC Bioinformatics. 2023;24(1):48. Tasci E et al. Bias and Class Imbalance in Oncologic Data-Towards Inclusive and Transferrable AI in Large Scale Oncology Data Sets. Cancers (Basel), 2022. 14(12). Tables Table 1 to 4 are available in the Supplementary Files section. Additional Declarations No competing interests reported. Supplementary Files Table1Patientscharacteristic18and24monthsMortality.docx Table2TheMostImportantVariable18and24monthmortality.docx Table3Characteristicsofpredictivepartitionanalyses.docx Table4LCAbasedconditionalprobability.docx Supplementarygastriccancerrevised.docx Cite Share Download PDF Status: Published Journal Publication published 30 Apr, 2025 Read the published version in BMC Cancer → Version 1 posted Editorial decision: Accepted 23 Apr, 2025 Editor assigned by journal 23 Apr, 2025 Reviews received at journal 03 Apr, 2025 Reviewers agreed at journal 02 Apr, 2025 Reviewers invited by journal 02 Apr, 2025 Submission checks completed at journal 27 Mar, 2025 First submitted to journal 27 Mar, 2025 You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-5731247","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":437712123,"identity":"260ed2cd-24ec-4913-8438-af07804760a9","order_by":0,"name":"Haniyeh Rafiepoor","email":"","orcid":"","institution":"Tehran University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Haniyeh","middleName":"","lastName":"Rafiepoor","suffix":""},{"id":437712124,"identity":"066fa691-0161-4504-9d7d-bb256a17577f","order_by":1,"name":"Mohammad Mehdi Banoei","email":"","orcid":"","institution":"University of Calgary","correspondingAuthor":false,"prefix":"","firstName":"Mohammad","middleName":"Mehdi","lastName":"Banoei","suffix":""},{"id":437712125,"identity":"364422a7-892d-4a76-b691-3c8f7fde03c4","order_by":2,"name":"Alireza Ghorbankhanloo","email":"","orcid":"","institution":"Tehran University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Alireza","middleName":"","lastName":"Ghorbankhanloo","suffix":""},{"id":437712126,"identity":"78af98f3-bd75-434d-b5ab-64f28e3f6a26","order_by":3,"name":"Ahad Muhammadnejad","email":"","orcid":"","institution":"Tehran University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Ahad","middleName":"","lastName":"Muhammadnejad","suffix":""},{"id":437712129,"identity":"7d7bd237-6e96-45c1-84f4-499ff613b185","order_by":4,"name":"Amirhossein Razavirad","email":"","orcid":"","institution":"Tehran University of Medical Sciences","correspondingAuthor":false,"prefix":"","firstName":"Amirhossein","middleName":"","lastName":"Razavirad","suffix":""},{"id":437712131,"identity":"2db75be1-52b4-4b66-a456-d0b01f709ba7","order_by":5,"name":"Saeed Soleymanjahi","email":"","orcid":"","institution":"Yale School of Medicine","correspondingAuthor":false,"prefix":"","firstName":"Saeed","middleName":"","lastName":"Soleymanjahi","suffix":""},{"id":437712132,"identity":"7c5b80cd-5477-42d9-8dd1-a219a205902a","order_by":6,"name":"Saeid Amanpour","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA3klEQVRIiWNgGAWjYHAD5gNw5gGcilABW2IDqVp4DBsIqgEB/mmHnz0uqDicz8B/5vtj3pw6Bv72A4yHK/BokbidZm4848xhywaJ3I3NvNsOM0icSWA4eAafNbcTzKR52w4bMEjwgrQAfXGDgeEgPifK307/BtHCf+YhUEsdgzwhLQa3c6C2MOQwArUwMxgQ0mJ4O6dMesaZdKDD0gxnzt12mMfwTGIDXi1yt9O3SRdUWAMddvjBh7fb6uTkjh8+/BGfFhBgBhH2ByAcHgYGRkIaoFpGwSgYBaNgFOAEAASCS4iFwYbgAAAAAElFTkSuQmCC","orcid":"","institution":"Tehran University of Medical Sciences","correspondingAuthor":true,"prefix":"","firstName":"Saeid","middleName":"","lastName":"Amanpour","suffix":""}],"badges":[],"createdAt":"2024-12-29 18:23:59","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-5731247/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-5731247/v1","draftVersion":[],"editorialEvents":[{"content":"https://doi.org/10.1186/s12885-025-14204-x","type":"published","date":"2025-04-30T15:57:15+00:00"}],"editorialNote":"","failedWorkflow":false,"files":[{"id":79906133,"identity":"c8a441af-e668-461c-a330-519533009d95","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":73594,"visible":true,"origin":"","legend":"\u003cp\u003eOverview of the study profile for AI application in gastric cancer prognosis. GC: gastric cancer, PPA: predictive partition analysis.\u003c/p\u003e","description":"","filename":"Figure1.flowchart.png","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/6077fbd3c2bcaf8d1bbe7577.png"},{"id":79906144,"identity":"a9f763a3-4232-4d28-9da0-3a6ba07bcd6e","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"jpg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":8617953,"visible":true,"origin":"","legend":"\u003cp\u003eSIMPLS-based scatter and coefficient plots show very good discrimination between A, B: survivors and non-survivors at 18 months and C, D: 24-monthmortality prediction studies.\u003c/p\u003e\n\u003cp\u003el Survivors, ¾ non-survivors.\u003c/p\u003e","description":"","filename":"Figure21824monthmortalityscatterandcoeffcientplots.edited.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/1fcc06abbe77412b23fc8560.jpg"},{"id":79906142,"identity":"5a91bf08-66b3-493b-82f3-e3a297d89478","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"jpg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":3531600,"visible":true,"origin":"","legend":"\u003cp\u003eA PCA-based correlation analysis heatmap revealed that the most differentiating variables were highly correlated with mortality at 18months in the A: non-survivor and B: survivor and 24 months in the C: non-survivor and D: survivor groups. Positive correlations are shown in red, whereas negative correlations are shown in blue. The intensity of the color is related to the correlation coefficient.\u003c/p\u003e","description":"","filename":"Figure3Correlationheatmapofallvariavleamongsurvivorsandnonsurvivors18and24month.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/db71af74ab4538dcfdaa50a6.jpg"},{"id":79906140,"identity":"0fb38994-da7f-4a2a-816c-d980650b0630","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":175357,"visible":true,"origin":"","legend":"\u003cp\u003eSee image above for figure legend\u003c/p\u003e","description":"","filename":"4.png","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/545e95fb468876d6bc53dd49.png"},{"id":79906148,"identity":"0f0fd2f9-8e3c-4970-9bfe-0daafae58e2c","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"jpg","order_by":5,"title":"Figure 5","display":"","copyAsset":false,"role":"figure","size":4867599,"visible":true,"origin":"","legend":"\u003cp\u003eCoefficient plots showing the most important variablesfor subtyping A: intestine B: adenocarcinoma and C: SIMPLS-based scatter plots showinggood discrimination between two categories of patients with TNM stage \u0026gt; IIb and ≤ IIb\u003c/p\u003e","description":"","filename":"Figure5CoefficientplotintestineandadenocarcinomaandTNMstagingscatterplot.jpg","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/54c948f40128ec0577d7c778.jpg"},{"id":81988439,"identity":"b8f911b0-c03c-4922-b2cb-a5ba46786d65","added_by":"auto","created_at":"2025-05-05 16:08:39","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":17986223,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/9a42ef58-59a3-4f93-a90b-3d25f823e914.pdf"},{"id":79906135,"identity":"b1782808-d74d-4deb-b9ba-edfb96f0a0d4","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"docx","order_by":1,"title":"","display":"","copyAsset":false,"role":"supplement","size":27445,"visible":true,"origin":"","legend":"","description":"","filename":"Table1Patientscharacteristic18and24monthsMortality.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/773ceadbb991a168504395c5.docx"},{"id":79908215,"identity":"9f114bbd-752f-4c2e-85fe-205fd904ec1f","added_by":"auto","created_at":"2025-04-04 11:18:10","extension":"docx","order_by":2,"title":"","display":"","copyAsset":false,"role":"supplement","size":14824,"visible":true,"origin":"","legend":"","description":"","filename":"Table2TheMostImportantVariable18and24monthmortality.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/ecc7c70261895577c5a268be.docx"},{"id":79906136,"identity":"97506b94-bcfa-4029-b4c9-7d05a8c3fb64","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"docx","order_by":3,"title":"","display":"","copyAsset":false,"role":"supplement","size":15063,"visible":true,"origin":"","legend":"","description":"","filename":"Table3Characteristicsofpredictivepartitionanalyses.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/5041d8af6691593daa6903fa.docx"},{"id":79906139,"identity":"26712233-2195-49ce-8f9d-fe2cb5ae9f63","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"docx","order_by":4,"title":"","display":"","copyAsset":false,"role":"supplement","size":16468,"visible":true,"origin":"","legend":"","description":"","filename":"Table4LCAbasedconditionalprobability.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/bb8c566c0611f43986eec1f5.docx"},{"id":79906147,"identity":"8eb27274-1a44-4a3c-8235-e148de32c444","added_by":"auto","created_at":"2025-04-04 11:02:10","extension":"docx","order_by":5,"title":"","display":"","copyAsset":false,"role":"supplement","size":2516731,"visible":true,"origin":"","legend":"","description":"","filename":"Supplementarygastriccancerrevised.docx","url":"https://assets-eu.researchsquare.com/files/rs-5731247/v1/56915b09810ef8bd356cba4c.docx"}],"financialInterests":"No competing interests reported.","formattedTitle":"Exploring the Potential of Machine Learning in Gastric Cancer: Prognostic Biomarkers, Subtyping, and Stratification","fulltext":[{"header":"Introduction","content":"\u003cp\u003eGastric cancer (GC), accounting for 7.7% of total cancer deaths, is the third leading cause of cancer deaths worldwide. According to estimates published by the International Agency for Research on Cancer (IARC), over 750,000 new mortalities of GC worldwide and more than 12,500 new cases of mortality have been reported in Iran. [\u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e1\u003c/span\u003e]\u003c/p\u003e \u003cp\u003eDespite all the advantages in the field of diagnosis and treatment of GC, the prognosis of patients with GC is still very poor, especially the long-term survival of patients with advanced GC, for whom the median survival is less than 12 months [\u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2\u003c/span\u003e]. Therefore, accurate staging of GCs in patients may improve the management and outcome of these patients [\u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e3\u003c/span\u003e]. The TNM staging system provided by the International Union against Cancer/American Joint Committee on Cancer (UICC/AJCC)) has been recognized as the most important staging system [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, the TNM staging system has limited applications, and middle-stage patients present diverse prognostic outcomes, making it difficult to accurately predict the mortality of patients. [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Advancements in the management of GC and innovative therapeutic approaches offer preoperative treatment options and highlight the importance of identifying high-risk patients and the role of biomarkers in GC management [\u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e6\u003c/span\u003e, \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e7\u003c/span\u003e]. These factors underscore the need for the development of new and accurate prognostic models [\u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e8\u003c/span\u003e, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e]. Using various artificial intelligence (AI) algorithms, nonlinear statistical models can be constructed to predict the survival of GC patients and categorize patients into groups with better distinguishing abilities via unsupervised machine learning (ML) algorithms [\u003cspan additionalcitationids=\"CR11 CR12\" citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the past few years, several models with different capabilities for determining the prognosis of GC have been designed [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan additionalcitationids=\"CR14 CR15 CR16\" citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Numbers of these models include only common clinical indicators (TNM stage, age, sex, etc.) to determine prognosis [\u003cspan additionalcitationids=\"CR15\" citationid=\"CR14\" class=\"CitationRef\"\u003e14\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR16\" class=\"CitationRef\"\u003e16\u003c/span\u003e], and few models consider biomarkers in addition to other clinical factors, which increases the accuracy of the models. However, the number of these biomarkers is limited and can be expanded [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e, \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e]. Additionally, the performed analysis has a limited ability, which can be solved by using AI and ML algorithms. Traditional analysis methods often have limitations in handling complex and multidimensional datasets, identifying nonlinear patterns, processing unstructured data, and performing predictive analysis. AI and ML algorithms can address these limitations and provide more robust and efficient solutions [\u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e18\u003c/span\u003e]. The study of multiple biomarkers and their interactions plays a crucial role in understanding the complex relationships among molecular pathways, particularly in diseases such as gastric cancer.\u003c/p\u003e \u003cp\u003eFurthermore, the potential interactions among biomarkers are often overlooked in conventional models, impeding the complex relationships among molecular pathways. Notably, the integration of molecular pathology and AI could lead to novel biomarkers that have diagnostic or prognostic value.\u003c/p\u003e \u003cp\u003eSeveral biomarkers, such as human epidermal growth factor receptor 2 (HER-2) [\u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e19\u003c/span\u003e], matrix metalloproteinase 7 (MMP-7) [\u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e20\u003c/span\u003e, \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e21\u003c/span\u003e], cyclooxygenase 2 (Cox2) [\u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e22\u003c/span\u003e, \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e23\u003c/span\u003e], vimentin [\u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e24\u003c/span\u003e, \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e], tumor protein 53 (TP53) [\u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e26\u003c/span\u003e], CD34 [\u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e27\u003c/span\u003e], and Ki67 [\u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e28\u003c/span\u003e], have shown potential as markers for angiogenesis and cell proliferation in the prognosis of patients with GC.\u003c/p\u003e \u003cp\u003eIn the present study, our objective was to utilize an AI model to identify potential biomarkers and their combinations for prognostic purposes and variant classifications in gastric cancer. An ML-based statistical method was applied to predict the mortality of GC patients and identify the complex interactions between predictors. The clustering method was used to design and classify the patients into different mortality groups (low-, moderate-, and high-risk groups) and to understand the role of each feature in the groups. To achieve this objective, a combination of seven biomarkers and other clinical and histopathological factors was employed in a cohort of GC patients. Identifying potential predictors is helpful for determining the appropriate prognosis and improving the understanding of the pathophysiological behavior of GC.\u003c/p\u003e"},{"header":"Methods and Materials","content":"\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003eData collection\u003c/h2\u003e \u003cp\u003eA retrospective cohort of 140 patients diagnosed with GC was included in the study, which was previously described by Razavirad et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], who traditionally analyzed the pathological impacts of biomarkers. Clinical data and histopathological findings from patients were collected at the Cancer Institute, Imam Khomeini Hospital Complex, Tehran University of Medical Sciences, I.R. of Iran, from April 2011 to January 2016. All patients who underwent neoadjuvant therapy were excluded from the study. The pathologic stage of patients was classified according to the 8th edition of the American Joint Committee on Cancer (AJCC) classification. Patients with stage I\u0026ndash;III disease (n\u0026thinsp;=\u0026thinsp;140) were analyzed, and the follow-up period was 60 months. A total of 21 variables, including patient demographics, histological findings and 7 biomarkers, were collected from pathologically confirmed patients with GC. Formalin-fixed, paraffin-embedded primary tumor specimens were retrieved from the Cancer Institute archives, sectioned into 3 \u0026micro;m slices, and processed for immunohistochemistry (IHC). Tissue sections were deparaffinized, antigen-retrieved in citrate buffer, blocked with bovine serum albumin, and incubated with specific antibodies against HER2, CD34, p53, Ki67, COX-2, MMP7, and vimentin. Immunoreactivity was assessed by two pathologists, with HER2 scored per CAP guidelines, CD34 evaluated for microvessel density, and p53, Ki67, COX-2, MMP7, and vimentin categorized based on staining intensity [\u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e30\u003c/span\u003e]. In this study, participants with missing values were excluded to minimize bias and ensure the accuracy of the ML models. Mortality prediction was performed at two time points, 18 and 24 months, as the median and mean survival times for GC patients, which is consistent with previous studies in Iran [\u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e31\u003c/span\u003e]. However, the current cohort was followed for 5 years, and only 4% of patients survived. There was a significant group imbalance between survivors and non-survivors from 24 months until 5 years. Because we have a relatively small sample size, group imbalance can lead to the development of biased classifiers, where the predictive performance is skewed toward the majority class, potentially resulting in accurate predictions for the minority class.\u003c/p\u003e \u003cp\u003eContinuous and ordinal variables were converted to binary variables via a categorical approach (cutoff point for each variable) using partition analysis (JMP pro, SAS). This study was approved by the ethics committee of Imam Khomeini Hospital Complex- Tehran University of Medical Sciences, Tehran, Iran (No. IR.TUMS.IKHC.REC.1400.001).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003eStatistical analysis\u003c/h2\u003e \u003cp\u003eThe statistically inspired modification of partial least squares (SIMPLS), an algorithm of PLS (a linear ML method) [\u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e32\u003c/span\u003e, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e] analysis, was applied to create a prediction model of 18- and 24-month mortality via patient demographics, histopathological findings and biomarker variables. SIMPLS-based prediction models were obtained through the training and validation sets. The most differentiating variables with variable importance in the projection (VIP)\u0026thinsp;\u0026gt;\u0026thinsp;1.0 were selected for the building prediction model. The validation set was randomly selected on the basis of 30% of the 140 patients with GC, which was also considered internal validation. Q\u003csup\u003e2\u003c/sup\u003e and R\u003csup\u003e2\u003c/sup\u003eY were counted as the goodness for predictability and the goodness of variability, respectively, to evaluate the performance of the SIMPLS-based prediction model. The best prediction models were obtained using the most differentiating variables when the Q\u003csup\u003e2\u003c/sup\u003e reached the highest value before decreasing and with the highest R\u003csup\u003e2\u003c/sup\u003eY. Q\u003csup\u003e2\u003c/sup\u003e and R\u003csup\u003e2\u003c/sup\u003eY were computed and verified through the training and validation sets, respectively. The performances of the models are evaluated via leave-one-out cross validation (LOOCV) and accuracy.\u003c/p\u003e \u003cp\u003eTo categorize the continuous and ordinal values, partition analysis was performed via a decision tree to find the best cutoff point of variables on the basis of the relationship between the outcome and predictors. Latent class analysis (LCA) was performed to identify low- and high-risk patients with GC through the clustering of cohorts using the most differentiating variables. Principal component analysis (PCA) was used to identify the different clusters obtained via LCA. PCA was also applied to find the trend and outliers using all the variables. Cox regression and Kaplan\u0026ndash;Meier survival analyses were performed on the AJCC stage groups (stages I, II, and III in the favorable category) among the test data and our model to compare the distinguishing ability between the two methods.\u003c/p\u003e \u003cp\u003eModel screening was conducted to check the final prediction model by providing a summary table via other ML methods, such as XGBoost, support vector machine (SVM), boosted tree (BT), bootstrap forest (BF), K-nearest neighbor (KNN), generalized regression lasso (GRL), boosting neural network (BNN), fit stepwise (FS), and na\u0026iuml;ve Bayes (NB) methods. Model screening also helps in finding efficient workflows as well as comparing and exploring datasets for the best predictive model.\u003c/p\u003e \u003cp\u003eWe further applied predictive partition analysis (PPA) or a decision tree, which is ML method [\u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e34\u003c/span\u003e] that could be used to classify or predict data responses. The ML aspect of the decision tree comes into determining when and where to divide the data between branches. Data partitioning was performed by dividing a dataset into subsets, including training and validation groups.\u003c/p\u003e \u003c/div\u003e"},{"header":"Results","content":"\u003cp\u003e\u003cstrong\u003ePatient\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;characteristics\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA total of 140 patients with confirmed GC were enrolled in the study, and 62 (44%) and 99 (70%) patients died within 18 and 24 months, respectively. Table 1 shows the demographic characteristics and histopathological and biochemical markers of survivors and non-survivors GC patients at 18 and 24 months after disease onset. The patient information is available in supplementary table E1.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePredicting Mortality in Gastric Cancer\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003ePatients via\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;Clinical and Paraclinical Data\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eAn overview of the present study is shown in Figure 1. The ML-based approach showed that demographics,\u0026nbsp;histopathological and biochemical markers can be used for predicting the mortality outcomes of patients with GC. SIMPLS analysis was carried out via most differentiating variables (VIPs) [35] to establish the prediction model. The prediction model was developed on 95 patients in the training set and 45 patients in the validation set. Two\u0026nbsp;factor-based SIMPLS models had high predictabilities\u0026nbsp;(Q\u003csup\u003e2\u0026nbsp;\u003c/sup\u003e= 0.45 and Q\u003csup\u003e2\u0026nbsp;\u003c/sup\u003e= 0.70)\u0026nbsp;for predicting\u0026nbsp;mortality\u0026nbsp;at\u0026nbsp;18 and 24\u0026nbsp;months, respectively,\u0026nbsp;and included\u0026nbsp;a total of 10 variables that contributed to the prediction models. SIMPLS-based scatter plots demonstrated very good discrimination between survivors and\u0026nbsp;non-survivors\u0026nbsp;for both 18- and 24-month mortality\u0026nbsp;prediction\u0026nbsp;studies (Figure 2. A \u0026amp; C). Table 2 shows that MMP-7, P53,\u0026nbsp;and Ki67\u0026nbsp;were the most important\u0026nbsp;variables for\u0026nbsp;18-month mortality.\u0026nbsp;For\u0026nbsp;24-month P53,\u0026nbsp;Ki67\u0026nbsp;and vimentin\u0026nbsp;were\u0026nbsp;considered the top 3 predictors associated\u0026nbsp;with\u0026nbsp;24-month\u0026nbsp;mortality in patients with GC.\u003c/p\u003e\n\u003cp\u003eFurthermore, the coefficient plots revealed that pathological tumor scoring and the presence of regional lymph nodes, vimentin, HER-2, COX, MMP-7, Ki67, and P53 were positively correlated with 18- and 24-month mortality. Low CD34 and high CD34 levels were correlated with mortality at 18- and 24-month mortality, respectively, indicating that the role of CD34 may change in mortality over time (Figure 2. B \u0026amp; D).\u003c/p\u003e\n\u003cp\u003ePPA verified that the\u0026nbsp;abovementioned most differentiating variables are strong predictors for mortality at 18 and 24 months for the training and validation sets in comparison with TNM staging (Table 3).\u003c/p\u003e\n\u003cp\u003ePCA-based correlation analysis showed that the\u0026nbsp;abovementioned differentiating variables were highly correlated with each other for 18- and 24-month mortality prediction (Figure 3). Age group was less correlated with most of the variables. Additionally, tumor size was positively correlated with regional lymph node variables in both 18- and 24-month mortality studies. PCA correlations among the most important variables\u0026nbsp;are available in supplementary table E2.\u003c/p\u003e\n\u003cp\u003eModel screening revealed high AUCs (\u0026gt; 0.80), high specificities (\u0026gt; 90%), and good sensitivities (\u0026gt;70%) when the most differentiating predictors were used in other ML methods, such as SVM, KNN, and GRL (Table E3).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eIdentification of high\u003c/strong\u003e\u003cstrong\u003e-\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;and low-risk\u0026nbsp;\u003c/strong\u003e\u003cstrong\u003epatients\u003c/strong\u003e\u003cstrong\u003e\u0026nbsp;with gastric cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eFurther investigations using LCA showed that patients with GC can be clustered to identify high-risk patients based on the clinical and biomarker data. LCA-based clustering revealed three main clusters among survivors and non-survivors. LCA-based clustering revealed that cluster 2 and cluster 3 had 44% and 70% mortality rates, respectively (Figure 4).\u0026nbsp;Compared with Clusters 2 and 3, Cluster 1 had the lowest rate of mortality (0%). All 3 clusters were well depicted through a PCA plot that can verify the clustering via two unsupervised methods. Table 4 shows that although variables had different contributions (conditional probabilities) to each cluster, several variables markedly impact clustering. Hence, CD34 \u0026ge; 30, P53 \u0026ge; 6, Ki67 \u0026ge; 13, and vimentin \u0026gt;2 were highly correlated with cluster 3 and the highest rate of mortality. On the other hand, the biomarkers MMP-7 \u0026le; 2, HER-2 \u0026le; 1, pathological tumor \u0026lt;3 and regional lymph node \u0026le; N1 were specifically correlated with the lowest rate of mortality (cluster 1). Moreover, HER-2 \u0026gt;1, regional lymph node \u0026gt; N1, COX-2 \u0026gt;2 and MMP-7 \u0026gt;2 showed similar probabilities for clusters 2 and 3. Cox regression analysis and Kaplan‒Meier survival curves revealed that the survival curve of each subgroup had good precision for 18- and 24-month mortality (\u003cem\u003ep\u0026nbsp;\u003c/em\u003e\u0026lt; 0.05). For both the clinical TNM staging and LCA clustering methods, the curves showed acceptable discrimination (\u003cem\u003ep\u003c/em\u003e \u0026lt; 0.05) (Supplementary Figure E1).\u003c/p\u003e\n\u003cp\u003eThe partition analysis based on 18-month mortality showed two main branches on two sides. The branches show that patients with P53 \u0026ge; 6, COX-2 \u0026gt;2 and vimentin \u0026gt;2 are associated with the prediction of mortality, and patients with P53 \u0026lt;6 and vimentin \u0026le; 2 are associated with survival (Supplementary Figure E2). Also, 24-month-based partition analysis showed that P53 \u0026ge; 6 and Ki67 \u0026ge; 13 in patients with GC were associated mainly with mortality. On the other hand, patients with P53 \u0026lt;6 and regional lymph nodes \u0026le; N1 were associated with survival outcomes (Supplementary Figure E3).\u003c/p\u003e\n\u003cp\u003eP\u003cstrong\u003eredicting TNM stages of gastric cancer\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA SIMPLS-based prediction model revealed that vimentin, P53 and HER-2 can predict the TNM stage of GC patients as an outcome with good predictability (Q2 = 0.45) (Figure 5C). These biomarkers were the most differentiating among demographics and histopathological and biochemical markers when used in dichotomized values. The predictive decision tree demonstrated cut-off points of vimentin (\u0026le; 2 or \u0026gt; 2), P53 (\u0026lt; 6 or \u0026ge; 6) and HER-2 (\u0026le; 1 or \u0026gt; 1) for the diagnosis of TNM-based staging, which was interestingly the same for the prediction of mortality outcome (Supplementary Figure E4). Model screening showed\u0026nbsp;high AUCs (\u0026gt; 0.80), high specificities (\u0026gt; 90%), and good sensitivities (\u0026gt;70%)\u0026nbsp;when\u0026nbsp;the most differentiating predictors\u0026nbsp;were used\u0026nbsp;in other ML methods,\u0026nbsp;such as SVM, KNN, and GRL (Supplementary Table E4). Decision tree analysis also\u0026nbsp;revealed\u0026nbsp;that vimentin is the most differentiating biomarker,\u0026nbsp;whereas vimentin \u0026lt;2 and vimentin \u0026gt;3 were mostly associated with TNM\u0026nbsp;stages\u0026nbsp;\u0026le; IIb and \u0026gt; IIb,\u0026nbsp;respectively (Supplementary Figure E4).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrediction of Different Types of Gastric Cancer: Tumor Location and Histology\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe applied the SIMPLS method to predict the tumor location (cardiac vs. non-cardiac) and tumor histology (intestine vs. non-intestine, diffuse vs. non-diffuse) as well as the tumor type (adenocarcinoma vs. non-adenocarcinoma) using clinical and biomarker variables. The models for differentiating adenocarcinoma from non-adenocarcinoma, and the intestine from the non-intestine were acceptably predictive (Q\u003csup\u003e2\u0026nbsp;\u003c/sup\u003e= 0.33 and Q\u003csup\u003e2\u0026nbsp;\u003c/sup\u003e= 0.34, respectively). Tumor histology, histology grade, tumor size, and age mostly contributed to predicting adenocarcinoma. Adenocarcinoma, pathological tumor, histology grade,\u0026nbsp;HER-2 status, and age contributed to the ability to predict the tumor histology of the intestine from that of the non-intestine. The coefficient plot indicates how variables are correlated with the GC intestine and non-intestine subtypes (Figure 5A-B).\u003c/p\u003e\n\u003cp\u003eMultivariate correlation analysis showed higher correlation between biomarkers in the cardia and intestine subtypes than between biomarkers in the noncardia and non-intestine subtypes. Age was negatively correlated with other variables among non-cardia subtypes compared with cardia. This phenomenon was also visible in the intestine and non-intestine subtypes (Supplementary Figure E5).\u003c/p\u003e"},{"header":"Discussion","content":"\u003cp\u003eDespite all the progress that has been made in the field of biomarker identification and model development in the field of oncology, there are still limitations in classifying patients in terms of prognosis and identifying the roles of various indicators and their relationships. In this study, the findings indicated that the ML-based SIMPLS model can accurately predict the mortality of GC patients via demographic and histopathological data, along with identifying the most important predictors and interactions between variables at two time points. The mean survival time for the patients was 18 months. The study focused on comparing the mortality rates between 18 and 24 months at 6-month intervals. In our study, 9 variables, including 2 clinical and 7 histopathological markers, were identified as potential prognostic predictors (Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e). Mortality predictor variables were weighed and ordered on the basis of their importance in the prediction model. Hence, for 18-month mortality, MMP-7\u0026thinsp;\u0026gt;\u0026thinsp;2, P53\u0026thinsp;\u0026ge;\u0026thinsp;6%, and Ki67\u0026thinsp;\u0026gt;\u0026thinsp;13 and for 24-month mortality, P53\u0026thinsp;\u0026ge;\u0026thinsp;6%, Ki67\u0026thinsp;\u0026gt;\u0026thinsp;13, and vimentin\u0026thinsp;\u0026gt;\u0026thinsp;2 were regarded as the top three predictors with the greatest impact on the model in the presence of other clinical risk factors, such as vascular invasion. This observation highlights the capacity of molecular biomarkers to provide a more accurate prognosis than do clinical factors. Nonetheless, the model demonstrated lower efficacy in the absence of additional predictors, as indicated in Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e.\u003c/p\u003e \u003cp\u003eFew numbers of studies have been conducted on the development of AI-based models in the field of GC prognosis. Some studies have utilized deep learning algorithms for image analysis [\u003cspan additionalcitationids=\"CR37\" citationid=\"CR36\" class=\"CitationRef\"\u003e36\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e38\u003c/span\u003e], whereas others have used numerical algorithms [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e, \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e10\u003c/span\u003e, \u003cspan additionalcitationids=\"CR13 CR14 CR15 CR16\" citationid=\"CR12\" class=\"CitationRef\"\u003e12\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e17\u003c/span\u003e, \u003cspan additionalcitationids=\"CR40\" citationid=\"CR39\" class=\"CitationRef\"\u003e39\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e41\u003c/span\u003e], most of them have used SVM and ANN algorithms for model development and reported accuracies ranging from 0.79\u0026ndash;0.94. Some of the studies included demographic (age, sex) and clinicopathological (tumor size, pathological grade, TNM staging) features and two biomarkers (CEA, CA199) as predictors. However, in the study of Jiang et al. [\u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e40\u003c/span\u003e], in addition to 3 baseline features (sex, CEA, and lymph node metastasis), 8 IHC markers (CD3invasive margin, CD3center of tumor, CD8IM, CD45ROCT, CD57IM, CD66bIM, CD68CT and CD34) were also considered. The main objective of most studies was to design a model using ML approaches to predict patient mortality with higher accuracy, which is traditionally predicted by TNM staging. TNM staging has been recognized as the most crucial staging system for stratifying GC patients into different risk groups [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e]. However, middle-stage patients have a variety of prognostic outcomes, and a more accurate classification of these patients is required [\u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e5\u003c/span\u003e]. Importantly, the TNM staging system is applicable only after surgery due to the requirement of pathological analysis of tumor specimens and examination of lymph nodes. Nevertheless, some studies have suggested new patient classification methods to improve the TNM staging system [\u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e4\u003c/span\u003e, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e42\u003c/span\u003e, \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e43\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the study of Razavirad et al. [\u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e29\u003c/span\u003e], the data were analyzed using traditional statistical analysis methods. The present study employed unsupervised ML techniques, PCA and LCA, to cluster patients. This approach differs from previous studies, such as those of Que SJ et al. [\u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e9\u003c/span\u003e] and Oh et al. [\u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e13\u003c/span\u003e], who utilized supervised ML algorithms to categorize patients into mortality groups based on predicted survival probability. In the unsupervised learning approach, unique classes can be identified, which might offer a superior classification method in comparison to previous approaches [\u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e44\u003c/span\u003e]. The importance of each feature in each cluster was also revealed through this unsupervised approach.\u003c/p\u003e \u003cp\u003eBy considering the incorporation of all the parameters in model development and a thorough discussion of the statistical process, this study has shed light on the importance of individual factors and their interplay with other variables in clinical and statistical aspects that were not addressed in previous studies. Given the complexity of AI analysis, it is beneficial to consider the associations between features and the formation of distinct clusters to enhance medical comprehensibility. Furthermore, identifying the interaction of parameters can offer insights into GC molecular pathways [\u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e45\u003c/span\u003e]. The correlations and interactions between different variables are presented in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, revealing significant links among biomarkers such as vimentin/CD34, Ki67/CD34, and p53/Ki67 in non-survival cases. In the survival groups, notable correlations were observed between vimentin/COX2 and MMP-7/p53, suggesting the possibility that similar molecular pathways contribute to these relationships. Additionally, alterations in correlations were detected during different periods for certain biomarkers, including HER-2/vimentin, HER-2/CD34, vimentin/Ki67, and Ki67/CD34.\u003c/p\u003e \u003cp\u003eSeveral studies have shown significant relationships between these markers, such as the coexistence of positive HER2 status and vimentin expression, indicating more aggressive proliferation of carcinoma cells [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. In the study of Ludovini et al. [\u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e47\u003c/span\u003e], the significant relationship between HER2 overexpression and CD34 has been identified as an important indicator reflecting microvessel density (MVD). A combination of p53/MMP-7 expression was found to be a prognostic indicator for Stage II/III GC. [\u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e48\u003c/span\u003e] The relationship between p53 expression and Ki-67 was not statistically significant, but the combination of p53 and COX-2 status was associated with significantly higher mean Ki-67 values in p53+/COX-2\u0026thinsp;+\u0026thinsp;tumors than in p53-/COX-2- tumors [\u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e49\u003c/span\u003e]. Elevated cytoplasmic expression of TSPAN1 was positively correlated with increased nuclear expression of Ki-67, accompanied by upregulation of MVD, as indicated by the stromal level of CD34 [\u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e50\u003c/span\u003e]. Vimentin immunolabeling in the tumor-associated stroma was positively correlated with Ki-67, and COX-2 was found to promote the production of MMP-2, MMP-9, VEGF, and vimentin while inhibiting E-cadherin production [\u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e51\u003c/span\u003e]. The relationship between Ki67 and p53 and the histological grade of GC indicates their importance in cell differentiation [\u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e52\u003c/span\u003e, \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e53\u003c/span\u003e], whereas the correlation between p53 and CD34 biomarkers and perineural invasion indicates the role of biomarkers in cancer invasion pathways [\u003cspan additionalcitationids=\"CR55 CR56\" citationid=\"CR54\" class=\"CitationRef\"\u003e54\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e57\u003c/span\u003e]. It has also been shown that Ki67, HER2, vimentin, MMP-7, p53 and CD34 are correlated with gastric cancer mortality, supporting previous works [\u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e25\u003c/span\u003e, \u003cspan additionalcitationids=\"CR59 CR60 CR61\" citationid=\"CR58\" class=\"CitationRef\"\u003e58\u003c/span\u003e\u0026ndash;\u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e62\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn contrast to previous studies, the patient prognosis in the current study was assessed at 2 time points, 18 and 24 months. The order of most of the parameters was similar; however, some features, such as MMP-7, differed in importance, which could indicate changes in the pathophysiological role of MMP-7 for 6 months. Differences in the pattern correlations between factors related to 18- and 24-month mortality are also notable for the variability in tumor behavior during this 6-month interval, which requires further investigation to obtain a better understanding of tumor behavior. The role of MMP-7 and other MMPs in tumor growth, invasion and spread can cause this effect [\u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e63\u003c/span\u003e, \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e64\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eIn the clinical setting, employing all biomarkers and features to predict patient mortality may be impractical. Consequently, prioritizing the significance of individual biomarkers to select a subset for prognostic purposes can offer a practical approach for accurate prediction. This approach facilitates the early identification of aggressive GC cases and may guide the implementation of different treatment strategies, such as more aggressive chemotherapy or immunotherapy. Also, by classifying patients into distinct risk categories, the study supports a move toward personalized medicine in the treatment of GC. Patients identified within high-risk clusters may benefit from closer monitoring and early intervention, while those in low-risk categories could potentially avoid unnecessary aggressive treatments.\u003c/p\u003e \u003cp\u003ePartition analysis revealed that specific biomarker thresholds, p53, COX-2, and pathological tumors for 18-month mortality, and p53, CD34, and regional lymph nodes for 24-month mortality, are indicative of accurate mortality prediction. The results of the partition analysis agreed with those of the SIMPLS prediction model, indicating the importance of the P53 biomarker in the prediction of mortality at different times. Decision tree analysis revealed some differences in the importance of some biomarkers in the prediction of mortality with respect to the VIP scores of biomarkers, which may be related to the difference in the algorithms used for the two methods. Nonetheless, we believe that SIMPLS can be the primary method for identifying the most differentiating variables, and a decision tree could be a secondary method with more applications in clinical settings. Clinical Decision Support (CDS) systems have previously played a significant role in health care informatics and could be further enhanced through the incorporation of decision trees and partition analysis. These methodologies provide clinically practical cutoffs for biomarker expression, making AI-driven prognostic models more interpretable for healthcare providers. Such models can be integrated into clinical decision-support systems, assisting in staging, treatment selection, and the prediction of survival outcomes. Furthermore, these predictive models can be embedded in pathological workflows to assist pathologists in identifying high-risk GC patients.\u003c/p\u003e \u003cp\u003eA similar analysis was performed, in which the TNM system was used to identify the most important factors depicted in supplementary figure E3, and interestingly, patients were classified into advanced and early groups according to the conventional definition of advanced and early GC [\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e65\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eWhile TNM staging relies on tumor size (T), lymph node involvement (N), and metastasis (M), our model integrates multiple biomarkers (P53, Ki67, vimentin, MMP-7, COX-2, HER-2, CD34) alongside clinical factors, achieving higher predictive accuracy (AUC: 0.84\u0026ndash;0.99) and better risk stratification. Unlike TNM staging, which is only applicable post-surgery, our model can provide preoperative prognostic information, guiding treatment decisions earlier. Additionally, decision tree analysis identified biomarker cutoff values (e.g., P53\u0026thinsp;\u0026ge;\u0026thinsp;6, Ki67\u0026thinsp;\u0026ge;\u0026thinsp;13) that can be used in clinical settings. However, while TNM staging is globally accepted and easy to interpret, our model requires external validation and AI explainability techniques before it can be more broadly adopted in clinical settings. Integrating ML-based approaches into clinical workflows could significantly enhance patient prognosis and personalized treatment strategies.\u003c/p\u003e \u003cp\u003ePrevious investigations discovered that HER-2 status was significantly associated with various clinical parameters, such as tumor depth or pathological tumor (pT) status, recurrence, distant metastasis, and positive lymph node (pN) status, which was supported by prior research [\u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e66\u003c/span\u003e]. Moderate to poorly differentiated carcinoma frequently shows P53-positive staining, which is often accompanied by lympho-vascular invasion. Moreover, an analysis based on the pT and pN stages revealed a significant rise in p53-positive cases [\u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e67\u003c/span\u003e]. The expression of vimentin was notably elevated in patients with advanced stages of GC, particularly those exhibiting poorly differentiated types of gastric carcinoma [\u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e46\u003c/span\u003e]. These findings suggest a potential role for these biomarkers in predicting clinical outcomes and guiding therapeutic decisions for patients with GC.\u003c/p\u003e \u003cp\u003eGiven the diverse environmental risk factors associated with GC subtypes and the recent development of molecular classifications such as The Cancer Genome Atlas (TCGA) [\u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e68\u003c/span\u003e] and Asian Cancer Research Group (ACRG) [\u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e69\u003c/span\u003e], it is imperative to comprehend the role of biomarkers in each anatomical (gastroesophageal junction/cardia and non-cardia) [\u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e70\u003c/span\u003e] and histological (Laur\u0026eacute;n classification) [\u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e71\u003c/span\u003e] subtype to gain insights into the pathophysiology of GC. This work offers an advantage over previous studies by investigating the differences in biomarker expression between GC subtypes. In intestinal and non-intestinal GC, variation in the expression of HER-2 has been revealed, as has greater expression in the diffuse subtype, which diverges from earlier findings [\u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e72\u003c/span\u003e]. In addition, as presented in supplementary figure E5, gastroesophageal junction/cardia GC shows different correlation patterns than non-cardia GC does, suggesting that these seven biomarkers contribute to a molecular pathway similar to that of non-cardia cancer.\u003c/p\u003e \u003cp\u003eThe present study is subject to certain constraints. First, the sample size is relatively small compared with that of prior investigations. Nevertheless, the SIMPLS model is considered to be a suitable algorithm for cohorts of limited size [\u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e33\u003c/span\u003e, \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e73\u003c/span\u003e]. On the other hand, recent research on various ML techniques has indicated that the effectiveness of ML-based methods is linked to the quality of the data, even with sample size as small as 40\u0026ndash;50 [\u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e74\u003c/span\u003e].\u003c/p\u003e \u003cp\u003eAdditionally, the major limitation of the current study is the lack of an external validation set, which makes the generalizability and robustness of our findings uncertain, as the results have only been assessed within the confines of the current dataset. Additionally, since all patients were from a single institution, potential selection bias may have occurred. Incorporating a large independent validation cohort would have strengthened the study by allowing us to confirm the reproducibility and applicability of our conclusions across different patient populations, thereby enhancing the overall reliability of our findings. The prediction model was developed by dividing the data into training and validation sets and internally validate using the LOOCV method. Further studies are required to repeat the process of data partitioning several times to obtain an average performance. Finally, owing to the low number of patients, it was not possible to stratify them into more than three categories for better patient classification. Despite the 5-year follow-up period, the survival rate of patients at the end of 60 months was only 4%, leading to a significant group imbalance between survivors and non-survivors from 24 months until 5 years. Since we have a relatively small sample size, group imbalance may give rise to the development of biased classifiers, where the predictive performance is skewed toward the majority class, potentially resulting in accurate predictions for the minority class [\u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e75\u003c/span\u003e]; therefore, two specific time points, 18 and 24 months as the median and mean survival times, were chosen for conducting a mortality analysis.\u003c/p\u003e \u003cp\u003eIn conclusion, the present study provides valuable insights into the use of ML-based techniques for prognostic biomarker identification in GC patients. Furthermore, the study classified patients into three categories and identified the major factors associated with mortality and staging in each group. These results indicate that molecular biomarkers could offer a more accurate prognosis compared to relying solely on clinical factors. Specifically, MMP-7, P53, Ki67, and vimentin were identified as the top predictors. The analysis of the correlation between various factors and their interactions revealed that the biomarker relationship pattern could differ among different anatomical and histopathological subtypes of GC, highlighting the possibility of distinct molecular pathways in GC pathogenesis. The partition analysis technique employed in this study offers a practical tool for identifying cutoffs and prioritizing markers to simplify models for clinical use. However, further research with large independent validation cohort is necessary to validate these findings and explore the potential of clinical and histopathological markers for enhancing prognostic accuracy in patients with GC.\u003c/p\u003e \u003cp\u003eFuture research should focus on validating the identified biomarkers in larger, independent cohorts to assess their reliability and clinical utility. Given that some of these biomarkers are associated with early tumor progression, their validation could contribute to the development of early screening programs, enabling the detection of GC at an earlier, more treatable stage. Additionally, the potential application of these biomarkers in blood-based tests should be explored for non-invasive monitoring of GC progression and recurrence through liquid biopsy approaches. Such advancements could enhance surveillance strategies and improve patient outcomes. Further investigation into the molecular pathways linked to these biomarkers may provide insights into GC pathogenesis and facilitate the development of targeted therapies. Moreover, conducting subgroup analyses based on other factors than histological type, such as age, sex in larger cohorts would enhance disease classification, improve understanding of GC pathophysiology, and refine risk stratification models for personalized treatment approaches.\u003c/p\u003e \u003cp\u003eOverall, this study contributes to the ongoing efforts to improve patient outcomes by advancing our comprehension of the intricate nature of cancer.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eAcknowledgments\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eWe would like to express our sincere gratitude to Dr. Kazem Zendehdel for his invaluable guidance and support throughout the preparation of this study. His expertise in the field of cancer epidemiology has been instrumental in shaping the direction of this work.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of data and materials\u0026nbsp;\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe data is available upon request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eA part of the analysis costs of this study was provided by a grant from the Cancer Biology Research Center, the Cancer Research Center, Tehran University of Medical Sciences (Grant No. 52012).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting of interest\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eEthics approval and consent to participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study was approved by the ethics committee of Imam Khomeini Hospital Complex- Tehran University of Medical Sciences, Tehran, Iran (No. IR.TUMS.IKHC.REC.1400.001). Written informed consent obtained from all the participants in the manuscript. All experiments were performed in accordance with the Guidelines of Iran National Committee for Ethics in Biomedical Research.\u0026nbsp;\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAuthor contributions\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eH.R. contributed to the study concept, data cleaning, statistical analysis interpretation and drafting of the manuscript. M.M.B. contributed to the statistical analysis interpretation, and drafting of the manuscript. A.G. contributed to the interpretation and drafting of the manuscript. A.M. contributed to the study design, data collection, patients follow up and preparation. A.R. contributed to the study concept, collection, patients follow up and data cleaning. S.S contributed to the data collection A.S.Z. contributed to statistical analysis and manuscript revision. S.A. contributed to interpretation and drafting of the manuscript, and overall supervision of the study. All authors read and approved the final manuscript.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003eBray F, et al. Global cancer statistics 2018: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2018;68(6):394\u0026ndash;424.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang XY, Zhang PY. Gastric cancer: somatic genetics as a guide to therapy. J Med Genet. 2017;54(5):305\u0026ndash;12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJanczewski LM, et al. The inaccuracies of gastric adenocarcinoma clinical staging and its predictive factors. J Surg Oncol. 2023;127(7):1116\u0026ndash;24.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang J, et al. Evaluation of the seventh AJCC TNM staging system for gastric cancer: a meta-analysis of cohort studies. Tumour Biol. 2014;35(9):8525\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu L, et al. The prognostic efficacy of the 8th edition UICC TNM classifications for gastric cancer in Chinese patients: A study based on follow-up system of nursing department. Med (Baltim). 2018;97(36):e12284.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCai H, et al. Advances in molecular biomarkers research and clinical application progress for gastric cancer immunotherapy. Biomark Res. 2022;10(1):67.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen Y, et al. Predicting response to immunotherapy in gastric cancer via multi-dimensional analyses of the tumour immune microenvironment. Nat Commun. 2022;13(1):4851.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAbbas M, et al. Current and future biomarkers in gastric cancer. Biomed Pharmacother. 2018;103:1688\u0026ndash;700.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eQue SJ, et al. Application of preoperative artificial neural network based on blood biomarkers and clinicopathological parameters for predicting long-term survival of patients with gastric cancer. World J Gastroenterol. 2019;25(43):6451\u0026ndash;64.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKorhani Kangi A, Bahrampour A. Predicting the Survival of Gastric Cancer Patients Using Artificial and Bayesian Neural Networks. Asian Pac J Cancer Prev. 2018;19(2):487\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLee J, et al. Deep Learning-Based Survival Analysis Identified Associations Between Molecular Subtype and Optimal Adjuvant Treatment of Patients With Gastric Cancer. JCO Clin Cancer Inf. 2018;2:1\u0026ndash;14.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eNilsaz-Dezfouli H, et al. Improving Gastric Cancer Outcome Prediction Using Single Time-Point Artificial Neural Network Models. Cancer Inf. 2017;16:1176935116686062.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOh SE, et al. Prediction of Overall Survival and Novel Classification of Patients with Gastric Cancer Using the Survival Recurrent Network. Ann Surg Oncol. 2018;25(5):1153\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eAmiri Z, et al. Assessing the effect of quantitative and qualitative predictors on gastric cancer individuals survival using hierarchical artificial neural network models. Iran Red Crescent Med J. 2013;15(1):42\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiglarian A, et al. Application of artificial neural network in predicting the survival rate of gastric cancer patients. Iran J Public Health. 2011;40(2):80\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBiglarian A, et al. Determining of prognostic factors in gastric cancer patients using artificial neural networks. Asian Pac J Cancer Prev. 2010;11(2):533\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhu L, et al. Comparison between artificial neural network and Cox regression model in predicting the survival rate of gastric cancer patients. Biomed Rep. 2013;1(5):757\u0026ndash;60.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang L, et al. From machine learning to deep learning: progress in machine intelligence for rational drug discovery. Drug Discov Today. 2017;22(11):1680\u0026ndash;5.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJ\u0026oslash;rgensen JT, Hersom M. HER2 as a Prognostic Marker in Gastric Cancer - A Systematic Analysis of Data from the Literature. J Cancer. 2012;3:137\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIi M, et al. Role of matrix metalloproteinase-7 (matrilysin) in human cancer invasion, apoptosis, growth, and angiogenesis. Exp Biol Med (Maywood). 2006;231(1):20\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKoskensalo S, et al. MMP-7 overexpression is an independent prognostic marker in gastric cancer. Tumour Biol. 2010;31(3):149\u0026ndash;55.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLim HY, et al. Increased expression of cyclooxygenase-2 protein in human gastric carcinoma. Clin Cancer Res. 2000;6(2):519\u0026ndash;25.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRistim\u0026auml;ki A, et al. Expression of cyclooxygenase-2 in human gastric carcinoma. Cancer Res. 1997;57(7):1276\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFuyuhiro Y, et al. Clinical significance of vimentin-positive gastric cancer cells. Anticancer Res. 2010;30(12):5239\u0026ndash;43.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOtsuki S, et al. Vimentin expression is associated with decreased survival in gastric cancer. Oncol Rep. 2011;25(5):1235\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei K, et al. The prognostic significance of p53 expression in gastric cancer: a meta-analysis. J Cancer Res Clin Oncol. 2015;141(4):735\u0026ndash;48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTenderenda M, et al. Expression of CD34 in gastric cancer and its correlation with histology, stage, proliferation activity, p53 expression and apoptotic index. Pathol Oncol Res. 2001;7(2):129\u0026ndash;34.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eXiong DD, et al. Ki-67/MKI67 as a Predictive Biomarker for Clinical Outcome in Gastric Cancer Patients: an Updated Meta-analysis and Systematic Review involving 53 Studies and 7078 Patients. J Cancer. 2019;10(22):5339\u0026ndash;54.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRazavirad A. Prognostic significance of HER2, p53, Ki67, MVD- CD34, Cox2, MMP7, and vimentin in gastric cancer. Basic Clin Cancer Res. 2023;15(2):67\u0026ndash;80.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePark YS et al. \u003cem\u003eA standardized pathology report for gastric cancer: 2nd edition.\u003c/em\u003e J Pathol Transl Med, 2023. 57(1): pp. 1\u0026ndash;27.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eFarahmandi F, et al. Survival of gastric cancer patients in Iran: a systematic review and meta-analysis. Gastroenterol Hepatol Bed Bench. 2023;16(3):245\u0026ndash;58.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBoulesteix AL, Strimmer K. Partial least squares: a versatile tool for the analysis of high-dimensional genomic data. Brief Bioinform. 2007;8(1):32\u0026ndash;44.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ede Jong S. SIMPLS: An alternative approach to partial least squares regression. Chemometr Intell Lab Syst. 1993;18(3):251\u0026ndash;63.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCharbuty B, Abdulazeez A. Classification Based on Decision Tree Algorithm for Machine Learning. J Appl Sci Technol Trends. 2021;2(01):20\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eEriksson L, et al. Multi- and Megavariate Data Analysis. Part I Basic Principles and Applications. Second revised and enlarged edition. Ume Sweden: MKS Umetrics AB; 2006. pp. 1\u0026ndash;103.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang Y, et al. Predicting peritoneal recurrence and disease-free survival from CT images in gastric cancer with multitask deep learning: a retrospective study. Lancet Digit Health. 2022;4(5):e340\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJin P, et al. Artificial intelligence in gastric cancer: a systematic review. J Cancer Res Clin Oncol. 2020;146(9):2339\u0026ndash;50.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWei T, et al. Survival prediction of stomach cancer using expression data and deep learning models with histopathological images. Cancer Sci. 2023;114(2):690\u0026ndash;701.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eIshii H, et al. Classification of gastric cancer subtypes using ICA, MLR and Bayesian network. Stud Health Technol Inf. 2013;192:1014.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJiang Y, et al. Immunomarker Support Vector Machine Classifier for Prediction of Gastric Cancer Survival and Adjuvant Chemotherapeutic Benefit. Clin Cancer Res. 2018;24(22):5574\u0026ndash;84.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhou C-M, et al. Predicting postoperative gastric cancer prognosis based on inflammatory factors and machine learning technology. BMC Med Inf Decis Mak. 2023;23(1):53.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen ZD, et al. Recent Advances in the Diagnosis, Staging, Treatment, and Prognosis of Advanced Gastric Cancer: A Literature Review. Front Med (Lausanne). 2021;8:744839.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLin JX et al. \u003cem\u003eMulticenter Validation Study of the American Joint Commission on Cancer (8th Edition) for Gastric Cancer: Proposal for a Simplified and Improved TNM Staging System.\u003c/em\u003e J Cancer, 2020. 11(12): pp. 3483\u0026ndash;3491.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLiu X. Supervised classification and unsupervised classification. ATS; 2005.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMoore JH, Raghavachari N. Artificial Intelligence Based Approaches to Identify Molecular Determinants of Exceptional Health and Life Span-An Interdisciplinary Workshop at the National Institute on Aging. Front Artif Intell. 2019;2:12.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhao W, et al. Clinical significance of vimentin expression and Her-2 status in patients with gastric carcinoma. Clin Transl Sci. 2013;6(3):184\u0026ndash;90.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLudovini V, et al. Evaluation of the prognostic role of vascular endothelial growth factor and microvessel density in stages I and II breast cancer patients. Breast Cancer Res Treat. 2003;81(2):159\u0026ndash;68.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSawada T, et al. New molecular staging with G-factor supplements TNM classification in gastric cancer: a multicenter collaborative research by the Japan Society for Gastroenterological Carcinogenesis G-Project committee. Gastric Cancer. 2015;18(1):119\u0026ndash;28.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eJoo YE, et al. Expression of cyclooxygenase-2, p53 and Ki-67 in gastric cancer. J Korean Med Sci. 2006;21(5):871\u0026ndash;6.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eChen L, et al. Clinicopathological significance of overexpression of TSPAN1, Ki67 and CD34 in gastric carcinoma. Tumori. 2008;94(4):531\u0026ndash;8.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eZhang X, et al. COX\u0026ndash;2 promotes epithelial\u0026ndash;mesenchymal transition and migration in osteosarcoma MG\u0026ndash;63 cells via PI3K/AKT/NF\u0026ndash;κB signaling. Mol Med Rep. 2019;20(4):3811\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLazăr D, et al. Ki-67 expression in gastric cancer. Results from a prospective study with long-term follow-up. Rom J Morphol Embryol. 2010;51(4):655\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLevine AJ, et al. The Role of the p53 Protein in Stem-Cell Biology and Epigenetic Regulation. Volume 6. Cold Spring Harb Perspect Med; 2016. 9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMuller PA, Vousden KH, Norman JC. \u003cem\u003ep53 and its mutants in tumor cell migration and invasion.\u003c/em\u003e J Cell Biol, 2011. 192(2): pp. 209\u0026thinsp;\u0026ndash;\u0026thinsp;18.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYang J, et al. Associations of CD34, Ki67, layer of invasion and clinical pathological characteristics, prognosis outcomes in gastrointestinal stromal tumors\u0026mdash;a retrospective cohort study. Translational Cancer Res. 2022;11(8):2866\u0026ndash;75.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBlanchet A et al. Isoforms of the p53 Family and Gastric Cancer: A M\u0026eacute;nage \u0026agrave; Trois for an Unfinished Affair. Cancers (Basel), 2021. 13(4).\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSaito H, et al. Neoangiogenesis and relationship to nuclear p53 accumulation and vascular endothelial growth factor expression in advanced gastric carcinoma. Oncology. 1999;57(2):164\u0026ndash;72.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHabibi A, et al. Correlation between CD34 Marker and Clinico-Pathologic Characteristics of Gastric Cancer. Int J Cancer Res. 2015;11:136\u0026ndash;42.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eKo GH, et al. Prognostic impact of Ki-67 in patients with gastric cancer-the importance of depth of invasion and histologic differentiation. Med (Baltim). 2017;96(25):e7181.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSoleyman-Jahi S, et al. Prognostic significance of matrix metalloproteinase-7 in gastric cancer survival: a meta-analysis. PLoS ONE. 2014;10(4):e0122316.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eUprak TK, et al. HER-2 incidence in gastric cancer, its association with prognosis and clinicopathological parameters. Ulus Cerrahi Derg. 2015;31(4):207\u0026ndash;13.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eYildirim M, et al. Prognostic significance of p53 in gastric cancer: a meta- analysis. Asian Pac J Cancer Prev. 2015;16(1):327\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eMurray GI, et al. Matrix metalloproteinases and their inhibitors in gastric cancer. Gut. 1998;43(6):791\u0026ndash;7.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003ePolistena A, et al. MMP7 expression in colorectal tumours of different stages. Vivo. 2014;28(1):105\u0026ndash;10.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eHu B, et al. Gastric cancer: Classification, histology and application of molecular pathology. J Gastrointest Oncol. 2012;3(3):251\u0026ndash;61.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eOgnjenovic L, et al. HER2 Positive Gastric Carcinomas and Their Clinico-Pathological Characteristics. Open Access Maced J Med Sci. 2018;6(7):1187\u0026ndash;92.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLazăr D, et al. The immunohistochemical expression of the p53-protein in gastric carcinomas. Correlation with clinicopathological factors and survival of patients. Rom J Morphol Embryol. 2010;51(2):249\u0026ndash;57.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eComprehensive molecular characterization of gastric adenocarcinoma. Nature, 2014. 513(7517): pp. 202\u0026ndash;9.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eCristescu R, et al. Molecular analysis of gastric cancer identifies subtypes associated with distinct clinical outcomes. Nat Med. 2015;21(5):449\u0026ndash;56.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eSiewert JR, et al. [Cardia cancer: attempt at a therapeutically relevant classification]. Chirurg. 1987;58(1):25\u0026ndash;32.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eLauren P, THE TWO HISTOLOGICAL MAIN TYPES. OF GASTRIC CARCINOMA: DIFFUSE AND SO-CALLED INTESTINAL-TYPE CARCINOMA. AN ATTEMPT AT A HISTO-CLINICAL CLASSIFICATION. Acta Pathol Microbiol Scand. 1965;64:31\u0026ndash;49.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eWang HB, Liao XF, Zhang J. Clinicopathological factors associated with HER2-positive gastric cancer: A meta-analysis. Med (Baltim). 2017;96(44):e8437.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eBanoei MM, et al. Unraveling complex relationships between COVID-19 risk factors using machine learning based models for predicting mortality of hospitalized patients and identification of high-risk group: a large retrospective study. Front Med (Lausanne). 2023;10:1170331.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eRajput D, Wang W-J, Chen C-C. Evaluation of a decided sample size in machine learning applications. BMC Bioinformatics. 2023;24(1):48.\u003c/span\u003e\u003c/li\u003e \u003cli\u003e\u003cspan\u003eTasci E et al. Bias and Class Imbalance in Oncologic Data-Towards Inclusive and Transferrable AI in Large Scale Oncology Data Sets. Cancers (Basel), 2022. 14(12).\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"},{"header":"Tables","content":"\u003cp\u003eTable 1 to 4 are available in the Supplementary Files section.\u003c/p\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":false,"hideJournal":false,"highlight":"","institution":"","isAcceptedByJournal":true,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true},"keywords":"Gastric cancer, Prediction model, Machine learning, Immunohistochemistry, Mortality","lastPublishedDoi":"10.21203/rs.3.rs-5731247/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-5731247/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003ch2\u003eBackground\u003c/h2\u003e \u003cp\u003eAdvancements in the management of gastric cancer (GC) and innovative therapeutic approaches highlight the significance of the role of biomarkers in GC prognosis. Machine-learning (ML)-based methods can be applied to identify the most important predictors and unravel their interactions to classify patients, which might guide prioritized treatment decisions.\u003c/p\u003e\u003ch2\u003eMethods\u003c/h2\u003e \u003cp\u003eA total of 140 patients with histopathological confirmed GC who underwent surgery between 2011 and 2016 were enrolled in the study. The inspired modification of the partial least squares (SIMPLS)-based model was used to identify the most significant predictors and interactions between variables. Predictive partition analysis was employed to establish the decision tree model to prioritize markers for clinical use. ML models have also been developed to predict TNM stage and different subtypes of GC. Latent class analysis (LCA) and principal component analysis (PCA) were carried out to cluster the GC patients and to find a subgroup of survivors who tended to die.\u003c/p\u003e\u003ch2\u003eResults\u003c/h2\u003e \u003cp\u003eThe findings revealed that the SIMPLS method was able to predict the mortality of GC patients with high predictabilities (Q\u003csup\u003e2\u003c/sup\u003e\u0026thinsp;=\u0026thinsp;0.45\u0026ndash;0.70). The analysis identified MMP-7, P53, Ki67, and vimentin as the top predictors. Correlation analysis revealed different patterns of prognostic markers in the non-survivor and survivor cohorts and different GC subtypes. The main prediction models were verified via other ML-based analyses, with a high area under the curve (AUC) (0.84\u0026ndash;0.99), specificity (0.82\u0026ndash;0.99) and sensitivity (0.87\u0026ndash;0.99). Patients were classified into three clusters of mortality risk, which highlighted the most significant mortality predictors. Partition analysis prioritizes the most significant predictors P53\u0026thinsp;\u0026ge;\u0026thinsp;6, COX-2\u0026thinsp;\u0026gt;\u0026thinsp;2, vimentin\u0026thinsp;\u0026gt;\u0026thinsp;2, Ki67\u0026thinsp;\u0026ge;\u0026thinsp;13 in mortality of patients (AUC\u0026thinsp;=\u0026thinsp;0.85\u0026ndash;0.90).\u003c/p\u003e\u003ch2\u003eConclusion\u003c/h2\u003e \u003cp\u003eThe present study highlights the importance of considering multiple variables and their interactions to predict the prognosis of mortality and stage in GC patients through ML-based techniques. These findings suggest that the incorporation of molecular biomarkers may enhance patient prognosis compared to relying solely on clinical factors. Furthermore, they demonstrate the potential for personalized medicine in GC treatment by identifying high-risk patients for early intervention and optimizing therapeutic strategies. The partition analysis technique offers a practical tool for identifying cutoffs and prioritizing markers for clinical application. Additionally, providing Clinical Decision Support systems with predictive tools can assist clinicians and pathologists in identifying aggressive cases, thereby improving patient outcomes while minimizing unnecessary treatments. Overall, this study contributes to the ongoing efforts to improve patient outcomes by advancing our comprehension of the intricate nature of GC.\u003c/p\u003e","manuscriptTitle":"Exploring the Potential of Machine Learning in Gastric Cancer: Prognostic Biomarkers, Subtyping, and Stratification","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-04-04 11:02:05","doi":"10.21203/rs.3.rs-5731247/v1","editorialEvents":[{"type":"communityComments","content":0},{"type":"decision","content":"Accepted","date":"2025-04-23T08:05:00+00:00","index":"","fulltext":""},{"type":"editorAssigned","content":"","date":"2025-04-23T08:04:25+00:00","index":"","fulltext":""},{"type":"editorInvitedReview","content":"","date":"2025-04-03T04:02:13+00:00","index":"hide","fulltext":""},{"type":"reviewerAgreed","content":"133983611406969307637795889731255310890","date":"2025-04-03T03:55:37+00:00","index":"hide","fulltext":""},{"type":"reviewersInvited","content":"","date":"2025-04-02T19:06:30+00:00","index":"","fulltext":""},{"type":"checksComplete","content":"","date":"2025-03-27T13:57:02+00:00","index":"","fulltext":""},{"type":"submitted","content":"BMC Cancer","date":"2025-03-27T08:55:45+00:00","index":"","fulltext":""}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"bmc-cancer","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":false,"externalIdentity":"bcan","sideBox":"Learn more about [BMC Cancer](http://bmccancer.biomedcentral.com/)","snPcode":"","submissionUrl":"https://www.editorialmanager.com/bcan/default.aspx","title":"BMC Cancer","twitterHandle":"BMC_series","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"em","reportingPortfolio":"BMC Series","inReviewEnabled":true,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"60421ca2-34bb-4cc0-9391-738bba7f6530","owner":[],"postedDate":"April 4th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"published-in-journal","subjectAreas":[],"tags":[],"updatedAt":"2025-05-05T16:06:27+00:00","versionOfRecord":{"articleIdentity":"rs-5731247","link":"https://doi.org/10.1186/s12885-025-14204-x","journal":{"identity":"bmc-cancer","isVorOnly":false,"title":"BMC Cancer"},"publishedOn":"2025-04-30 15:57:15","publishedOnDateReadable":"April 30th, 2025"},"versionCreatedAt":"2025-04-04 11:02:05","video":"","vorDoi":"10.1186/s12885-025-14204-x","vorDoiUrl":"https://doi.org/10.1186/s12885-025-14204-x","workflowStages":[]},"version":"v1","identity":"rs-5731247","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-5731247","identity":"rs-5731247","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}