Abstract
Purpose
Peritoneal metastasis from gastric cancer (GC) is generally considered incurable. Although conversion surgery is the preferred strategy, surgeons frequently encounter clinical scenarios requiring upfront surgery. Currently, objective criteria for predicting prognosis after upfront R0 gastrectomy remain undefined.
Materials and Methods
This retrospective multicenter cohort study analyzed data from 792 patients derived from the PASS-META cohort and an independent validation dataset collected between 2014 and 2021. The final analysis included 80 patients with peritoneal oligometastasis (P1/P2) who underwent upfront R0 gastrectomy for risk stratification, and 632 controls who underwent R2 gastrectomy or no gastrectomy for survival comparison. A component-wise gradient-boosting survival model was developed and simplified into an 8-item risk scoring system.
Results
The machine learning (ML) model demonstrated robust discrimination in external validation (concordance index, 0.811; 95% confidence interval, 0.677–0.955) and stratified patients into low- and high-risk groups with significantly different survival outcomes (log-rank P<0.001), independent of systemic chemotherapy compliance. Low-risk patients achieved a 5-year survival rate exceeding 30%, whereas high-risk patients showed survival rates comparable to those of patients who underwent R2 gastrectomy or no gastrectomy. The simplified 8-item risk groups showed high concordance with the ML model risk groups (F1-score, 0.975).
Conclusions
We developed and validated a prediction model and an 8-item risk scoring system that effectively stratified patients with peritoneal oligometastatic GC according to their prognosis after upfront R0 gastrectomy. Although prospective validation is warranted, this framework provides objective prognostic guidance that may support intraoperative decision-making, distinguishing patients with favorable long-term outcomes from those whose survival is comparable to that of noncurative approaches.
Keywords: Peritoneal oligometastasis, Upfront R0 gastrectomy, Machine learning
INTRODUCTION
Peritoneal metastasis in gastric cancer (GC) has traditionally been considered a terminal condition with a poor prognosis because of its resistance to systemic therapy and the lack of survival benefit from surgical resection. Currently, although systemic chemotherapy is initiated with a palliative aim, conversion surgery is regarded as the preferred strategy if a favorable chemotherapy response is observed [1,2,3,4]. However, in real-world practice, this approach is not always feasible because urgent surgery may be required for tumor-related symptoms, such as bleeding or obstruction [5]. Therefore, rather than uniformly applying a conversion surgery strategy, individualized therapeutic strategies tailored to each patient’s clinical context are essential. However, this need has not been sufficiently addressed in the management of patients with GC and peritoneal metastasis. To address this unmet clinical need, our recent large-scale multicenter study (PASS-META) provided compelling evidence for the oncologic value of gastrectomy [6]. By analyzing 983 patients with surgically confirmed stage IV GC, we demonstrated that gastrectomy was associated with significantly improved overall survival (OS) among patients with limited peritoneal metastases (P1 and P2). This finding suggests that, when surgical intervention is feasible, it can offer a distinct survival advantage in selected patient subgroups, thereby challenging the traditional chemotherapy-only paradigm [7]. However, translating these population-level findings into precision surgery requires further investigation. Although the P1/P2 subgroup generally shows a survival benefit after gastrectomy, the optimal criteria for prognostic stratification within this population remain undefined. Reliance solely on technical resectability may not be sufficient to ensure a favorable outcome for every patient. Therefore, investigation of whether specific clinical prognostic factors can further refine patient stratification for upfront R0 gastrectomy is necessary.
Therefore, the aim of this study was to apply a machine learning (ML)-based survival algorithm to a multi-institutional database to stratify patients with GC and peritoneal oligometastasis according to their prognosis after curative-intent R0 gastrectomy. Specifically, we sought to distinguish patients with favorable long-term survival from those whose outcomes were comparable to those associated with noncurative approaches. Furthermore, if patients can be accurately stratified using preoperative and intraoperative clinical information, surgical decision-making can be better informed, facilitating timely intervention for patients with favorable prognostic profiles. Here, we present the development and validation of this prognostic model and an accompanying simple risk-scoring system, providing a practical, data-driven framework to guide critical surgical decisions.
MATERIALS AND METHODS
Data collection
Patient data were retrospectively collected from a multicenter database of patients with pathologically confirmed peritoneal metastasis from GC treated at 6 institutions in South Korea (Ajou University Hospital, Asan Medical Center, Chonnam National University Hospital, Seoul St. Mary’s Hospital, Severance Hospital, and Dankook University Hospital) between 2014 and 2021. A total of 792 patients with GC and synchronous peritoneal metastasis, but without distant organ or distant lymph node metastasis, were included. The extent of peritoneal metastasis was classified as P1, P2, or P3 according to the Japanese Research Society for Gastric Cancer classification (Supplementary Fig. 1) [8], which has been applied consistently in both previous and current studies [9]. The study protocol was approved by the Institutional Review Board of Severance Hospital, Yonsei University Health System (4-2025-0144), which waived the requirement for informed consent because of the retrospective nature of the study.
Feature selection
To ensure immediate clinical utility, the ML model incorporated readily available preoperative and intraoperative variables. Selected features included demographics (age, sex, and body mass index), preoperative laboratory findings (hemoglobin, platelet, lymphocyte, neutrophil counts, albumin, and carbohydrate antigen 19-9 [CA19-9]); American Society of Anesthesiologists (ASA) physical status classification, tumor characteristics (size, location, Borrmann type, and differentiation), preoperative presentations (bleeding, obstruction, and perforation), and intraoperative findings (clinical T/N stages and Krukenberg tumor).
Development of the ML survival model
A component-wise gradient-boosting survival analysis (GBSA) model was developed to predict OS using the scikit-survival library in Python (Python Software Foundation, Beaverton, OR, USA) [10]. The GBSA model was selected as the primary modeling approach because it accommodates nonlinear predictor–outcome relationships, does not require the proportional hazards assumption, and performs stable component-wise variable selection with small to moderate sample sizes. The Cox proportional hazards model was prespecified as the benchmark comparator, and the performance of the two models was compared using the concordance index in the external validation cohort, with uncertainty quantified using 10,000 bootstrap resamples. To evaluate the model’s generalizability, we performed geographic external validation in accordance with the Transparent Reporting of a Multivariable Prediction Model for Individual Prognosis or Diagnosis framework [11]. The training dataset included data from four institutions (Ajou University Hospital, Asan Medical Center, Chonnam National University Hospital, and Seoul St. Mary’s Hospital), whereas the validation dataset consisted of data from 2 geographically independent institutions (Severance Hospital and Dankook University Hospital) with separate surgical teams and referral networks. Although all 6 institutions participated in the PASS-META consortium, nonrandom institution-based partitioning ensured that the development and validation sets differed in case mix and clinical practice patterns rather than by chance alone. To optimize the model’s hyperparameters (n_estimators and learning_rate) and reduce overfitting given the limited sample size, we implemented a bootstrap-based tuning strategy. Shuffle-split cross-validation (100 iterations with a 30% out-of-bag set) was applied to the training data, and the hyperparameter combination that achieved the highest mean concordance index (C-index) on the out-of-bag samples was selected. The final model was then retrained on the entire training dataset using the optimized parameters.
Model performance evaluation and risk stratification
The predictive performance of the model was evaluated in both the training and external validation datasets using the C-index. The final GBSA model was used to calculate continuous risk scores for each patient. To stratify patients into risk groups, a receiver operating characteristic (ROC) curve was generated using the 2-year OS status as the binary classification endpoint and risk scores from the training set (n=62; one patient who was alive and censored before 24 months was excluded). The optimal cutoff value was determined by maximizing Youden’s J-statistic. This cutoff was then applied to classify all patients into low- and high-risk groups. In addition, calibration was assessed at the 2-year time point in the training cohort by stratifying patients into quartiles of predicted risk and comparing the mean predicted 2-year survival probability with the observed Kaplan–Meier estimate within each quartile.
Survival analysis
For survival analysis of patients with peritoneal oligometastatic GC who underwent R0 gastrectomy, three endpoints were evaluated: OS, peritoneal progression-free survival (PFS), and OS after peritoneal progression, defined as the survival period from the time of peritoneal recurrence following the initial R0 gastrectomy. All survival analyses, including Kaplan–Meier curve generation and log-rank testing, were performed using the Lifelines library (version 0.27.8).
Feature importance analysis
We used 2 complementary approaches to interpret feature contributions. First, permutation-based feature importance was calculated for the validation set by measuring the average decrease in the C-index after randomly shuffling each feature (30 iterations). Second, internal model coefficients were analyzed to determine the direction and magnitude of the risk association; positive coefficients indicated an adverse prognosis, whereas negative coefficients indicated a favorable prognosis.
Development of a simple 8-item risk scoring system
To enhance clinical usability, a simplified integer-based risk scoring system was developed using the GBSA model. The features used for this scoring system were selected from the predictors constituting the GBSA model, excluding preoperative CA19-9, which was the only continuous variable and ranked lowest in terms of both permutation importance and coefficient magnitude. The remaining 5 favorable and three adverse predictors were selected based on the absolute magnitudes of their model coefficients. Each adverse factor was assigned +1 point, and each favorable factor was assigned −1 point. A total “simple score” was then calculated for each patient by summing these values. Patients were classified into 2 categories: the simple-high (score ≥0) and simple-low (score <0) risk groups. The ability of this simplified scoring system to replicate the stratification of the original GBSA model was evaluated using a confusion matrix, with agreement assessed using standard classification metrics, including accuracy and the macro-averaged F1-score. This simplified score is not an independent predictive model but a clinical approximation of the GBSA model.
Statistical analysis
All statistical analyses were performed using Python (version 3.8.12; Python Software Foundation). Comparisons of categorical and continuous variables were performed using the SciPy (version 1.10.1) and scikit-learn (version 1.0.2) libraries. Statistical significance was defined as a 2-sided P-value of <0.05.
RESULTS
Whole cohort, gastrectomy vs. no gastrectomy
Among the 792 patients, the gastrectomy group (n=221) showed significantly improved OS compared with the non-gastrectomy group (n=571) (log-rank P<0.001; Fig. 1, Supplementary Fig. 2A). However, this comparison was substantially confounded by selection bias, as the gastrectomy group was significantly enriched for low-burden (P1) cases associated with an inherently better prognosis (χ2 test, P<0.001; Supplementary Fig. 2B and C). To eliminate this confounding factor and isolate the true oncological impact of surgery, subsequent analyses focused exclusively on the gastrectomy cohort.
Fig. 1. Flow diagram of the study cohort.
Flowchart illustrating the selection of the patient group for this analysis.
GC = gastric cancer.
Gastrectomy cohort, R0 gastrectomy vs. R2 gastrectomy in patients with P1/2/3
In the gastrectomy group (n=221), potentially curative R0 resection demonstrated a significant survival benefit over R2 resection (log-rank P=0.001; Fig. 1, Supplementary Fig. 3A). However, stratified analysis revealed that patients with P3 had consistently poor outcomes regardless of resection status (Supplementary Fig. 3B) and were disproportionately overrepresented in the R2 group (χ2 test, P<0.001; Supplementary Fig. 3C). This indicated that the survival advantage of R0 resection in the gastrectomy group was substantially confounded by the high tumor burden of P3 disease.
Gastrectomy cohort, R0 gastrectomy vs. R2 gastrectomy in patients with P1/2
To isolate the oncological effects of R0 gastrectomy, we limited the cohort to patients with P1/P2 disease in the gastrectomy group (Fig. 1). In this subset, R0 gastrectomy showed a favorable trend compared with R2 gastrectomy but did not reach statistical significance (log-rank P=0.072; Supplementary Fig. 4A). However, significant survival benefits were observed in specific subgroups, such as Borrmann type I/II tumors (log-rank P=0.038; Supplementary Fig. 4B) and preoperative bleeding (log-rank P=0.011; Supplementary Fig. 4C). The observed heterogeneity in survival outcomes underscores the potential value of prognostic stratification based on clinical characteristics. In addition, the baseline characteristics of the entire cohort, stratified by treatment group and peritoneal metastasis stage, are summarized in Supplementary Table 1.
Development of a ML model predicting survival
To stratify patients according to their prognosis after R0 gastrectomy, we developed a GBSA model within the R0 gastrectomy group (Fig. 2A, Methods). The training and external validation cohorts, whose baseline characteristics are compared in Supplementary Table 2, were derived from geographically independent institutions (Methods). Following a bootstrap-based hyperparameter tuning process (Supplementary Fig. 5A, Methods), the final model demonstrated robust performance, achieving a C-index of 0.690 in the training set and 0.811 (bootstrapped 95% confidence interval [CI], 0.677–0.955; 13 events among 17 patients) in the external validation set (Fig. 2B, Supplementary Fig. 5B). To compare the GBSA model with the conventional Cox proportional hazards model, we fitted both models using the same feature set. The GBSA model demonstrated superior discrimination in the external validation cohort (C-index, 0.811 vs. 0.758; Supplementary Fig. 5C). In the bootstrapped analysis (10,000 iterations), the GBSA model outperformed the Cox model in 83.0% of the resampled datasets (mean delta C-index, 0.056; 95% CI, −0.041 to 0.173) (Supplementary Fig. 5C). The model generated a continuous risk score for each patient, ranging from −0.8711 to 2.235 (Fig. 2C), with lower scores indicating a more favorable prognosis and higher scores indicated a poorer prognosis. To confirm the clinical relevance of the predicted risk scores, we assessed their correlation with observed survival times. A significant, moderate negative correlation was observed (Pearson’s r=−0.51, P<0.001; Fig. 2D), demonstrating that higher predicted risk scores were strongly associated with shorter OS.
Fig. 2. Development of the ML survival model.
(A) Schematic overview of the study design for ML survival model development, including patient selection, dataset splitting, algorithm application, and model evaluation. (B) Bar plot showing the performance of the final GBSA model, as measured by the C-index, on the training and validation sets. (C) Histogram illustrating the distribution of the continuous risk scores generated by the ML model for the 80 patients in the analysis cohort. (D) Scatter plot demonstrating a significant negative correlation between the model-predicted risk score and observed OS time. Censored and death events are distinguished by white and black dots, respectively, with the Pearson’s correlation coefficient (r) and P-value indicated on the plot.
ML = machine learning; GBSA = gradient-boosting survival analysis; C-index = concordance index; OS = overall survival; GC = gastric cancer.
Machine learning risk score-based survival group stratification and prognostic implications
Using an optimized cutoff determined by ROC analysis (area under the curve [AUC]=0.798; Fig. 3A), the model classified patients into low- and high-risk groups (Table 1). Time-dependent AUC analysis demonstrated consistent discriminative performance across multiple time points (12 months, 0.727; 24 months, 0.798; 36 months, 0.728; 48 months, 0.774; and 60 months, 0.814), confirming that the model’s discrimination was not dependent on a specific time horizon (Supplementary Fig. 6A). Calibration analysis at the 2-year time point demonstrated a monotonically increasing relationship between the predicted and observed survival probabilities across risk quartiles, indicating that the model correctly ranked patients by risk (Supplementary Fig. 6B). This stratification demonstrated a significant survival difference in the validation set (log-rank P<0.001; Fig. 3B), which was further confirmed in the entire R0 gastrectomy group (log-rank P<0.001; Supplementary Fig. 7A). Beyond OS, the low-risk group showed significantly increased peritoneal PFS (log-rank P=0.001; Fig. 3C) and postprogression survival (log-rank P<0.001; Supplementary Fig. 7B), suggesting that the model captures intrinsic biological aggressiveness. No significant differences in systemic chemotherapy compliance were observed between the risk groups in terms of total cycles (t-test, P=0.841) or administration rates (χ2 test, P=0.419), ruling out treatment disparity as a confounder (Supplementary Fig. 8). Finally, we compared the survival outcomes across surgical strategies. High-risk patients showed survival rates comparable to those of patients who underwent R2 gastrectomy or no gastrectomy (Fig. 3D). In contrast, low-risk patients achieved significantly longer survival than the control groups (log-rank P<0.001; Fig. 3D). These results confirmed that the model effectively distinguished prognostic subgroups with distinct survival outcomes after R0 gastrectomy.
Fig. 3. Risk stratification using the ML model and its prognostic validation.
(A) ROC curve analysis using 2-year OS status as the binary classification endpoint on the training dataset, used to determine the optimal risk score cutoff for patient stratification. Kaplan-Meier curves demonstrating the prognostic significance of the risk stratification for (B) OS in the validation set (n=17) and (C) peritoneal PFS in the R0 gastrectomy group (n=80). (D) Kaplan-Meier curves for OS comparing the ML model-defined low-risk group against other patient groups, including the ML-defined high-risk group, R2 gastrectomy group, and no gastrectomy group. P-values were calculated by the log-rank test.
ML = machine learning; ROC = receiver operating characteristic; OS = overall survival; PFS = progression free survival; AUC = area under the curve.
Table 1. Baseline characteristics of patients who underwent R0 gastrectomy according to risk group.
| Variable | High-risk (n=43) | Low-risk (n=37) | P-value | |
|---|---|---|---|---|
| Age (yr) | 62.6±14.2 | 59.4±11.6 | 0.277 | |
| Sex, male | 26 (60.5) | 21 (56.8) | 0.914 | |
| BMI (kg/m2) | 22.2±2.9 | 22.8±2.9 | 0.355 | |
| ASA ≥III | 10 (23.3) | 0 (0.0) | 0.005 | |
| Preoperative bleeding | 1 (2.3) | 5 (13.5) | 0.142 | |
| Preoperative obstruction | 11 (25.6) | 1 (2.7) | 0.011 | |
| Preoperative perforation | 2 (4.7) | 0 (0.0) | 0.542 | |
| Histology | 0.060 | |||
| Well/moderately differentiated | 3 (7.0) | 11 (29.7) | ||
| Poorly differentiated | 18 (41.9) | 15 (40.5) | ||
| Signet ring cell | 20 (46.5) | 11 (29.7) | ||
| Unknown | 2 (4.7) | 0 (0.0) | ||
| Borrmann type | < 0.001 | |||
| I/II | 0 (0.0) | 12 (32.4) | ||
| III | 34 (79.1) | 14 (37.8) | ||
| IV | 8 (18.6) | 9 (24.3) | ||
| Unknown | 1 (2.3) | 2 (5.4) | ||
| Upper body involvement | 8 (18.6) | 21 (56.8) | 0.001 | |
| sT stage | 0.212 | |||
| Serosa exposure | 34 (79.1) | 23 (62.2) | ||
| Serosa invasion | 7 (16.3) | 13 (35.1) | ||
| Subserosa | 1 (2.3) | 1 (2.7) | ||
| Unknown | 1 (2.3) | 0 (0.0) | ||
| sN stage | 0.481 | |||
| N0 | 1 (2.3) | 0 (0.0) | ||
| N1 | 11 (25.6) | 7 (18.9) | ||
| N2 | 31 (72.1) | 30 (81.1) | ||
| Peritoneal metastasis stage | 1.000 | |||
| P1 | 38 (88.4) | 33 (89.2) | ||
| P2 | 5 (11.6) | 4 (10.8) | ||
| Krukenberg tumor | 0 (0.0) | 2 (5.4) | 0.409 | |
| Preoperative laboratory values | ||||
| Hemoglobin (g/dL) | 12.2±2.3 | 12.3±2.5 | 0.804 | |
| Platelet (×103/µL) | 278.2±90.9 | 288.7±87.3 | 0.602 | |
| Albumin (g/dL) | 4.0±0.5 | 4.0±0.6 | 0.888 | |
| CA19-9 (U/mL) | 113.6±331.3 | 78.9±393.0 | 0.669 | |
| CRP (mg/dL) | 1.2±2.6 | 0.2±0.5 | 0.012 | |
| Neutrophil (%) | 49.3±27.4 | 43.0±26.5 | 0.307 | |
| Lymphocyte (%) | 21.5±14.7 | 23.1±15.6 | 0.648 | |
| Postoperative chemotherapy | 0.419 | |||
| Yes | 38 (88.4) | 32 (86.5) | ||
| No | 5 (11.6) | 5 (13.5) | ||
| Chemotherapy cycles | 8.6±6.9 | 8.9±6.6 | 0.841 | |
Values are presented as mean ± standard deviation or number (%). P-values were calculated using Student’s t-test for continuous variables and Fisher’s exact test or χ2 test for categorical variables.
BMI = body mass index; ASA = American Society of Anesthesiologists; sT stage = surgical T stage; sN stage = surgical N stage; CA19-9 = carbohydrate antigen 19-9; CRP = C-reactive protein.
Feature importance analysis
To identify the key prognostic features, we used two complementary approaches. Permutation importance quantified the overall contribution of each feature by measuring the predictive loss when the feature was randomly shuffled. Complementing this, the model coefficients indicated the direction of association, with positive values representing adverse risk and negative values representing a favorable prognosis (Methods). Permutation analysis identified the most influential variables, including ASA III or higher, Krukenberg tumor, and Borrmann type I/II (Fig. 4A). The model coefficients consistently clarified their roles. The adverse factors included preoperative perforation, ASA physical status class III or higher, obstruction, and elevated CA19-9 levels (Fig. 4B). Conversely, favorable predictors included Krukenberg tumor, preoperative bleeding, well/moderate differentiation, Borrmann type I/II tumors, and upper body location. The strong correlation between permutation scores and absolute coefficient values (Spearman’s R=0.86, P=0.019; Fig. 4C) confirmed that the model reliably captured both the magnitude and direction of each clinical factor.
Fig. 4. Feature importance analysis of the GBSA model.
(A) Permutation-based feature importance calculated on the validation set. The length of the bar represents the overall impact of each feature on the model’s predictive performance. (B) Coefficients from the final trained GBSA model. A positive coefficient indicates an association with a poor prognosis (higher risk), while a negative coefficient indicates a favorable prognosis. (C) Scatter plot showing the correlation between the absolute values of the model coefficients and the permutation importance scores. The Spearman correlation coefficient (R) and P-value are indicated, and the correlation trendline is displayed as a dotted line.
GBSA = gradient-boosting survival analysis; ASA = American Society of Anesthesiologists; CA19-9 = carbohydrate antigen 19-9; Pre op = preoperative.
The simple 8-item risk scoring system for clinical application
To enhance the clinical usability of the computationally derived continuous risk score from the GBSA model, we developed a simplified integer-based risk scoring system that can be easily calculated in clinical practice. To construct this simple scoring system, we selected eight categorical features from the model, excluding preoperative CA19-9, which was the only continuous variable among the important features (Fig. 5A). The selected features consisted of 5 favorable predictors (negative coefficients): Krukenberg tumor, preoperative bleeding, well- or moderately differentiated histology, Borrmann type I or II tumor, and upper-body tumor location; and three adverse predictors (positive coefficients): preoperative obstruction, ASA physical status class III or higher, and preoperative gastric perforation. Each adverse predictor was assigned +1 point, and each favorable predictor was assigned −1 point. The total “simple score” was then calculated as the sum of all applicable feature scores for each patient. The resulting scores ranged from −3 to +2 (Supplementary Fig. 9A). To determine a clinically meaningful cutoff point, we analyzed OS according to each score within the R0 gastrectomy cohort. A significant decrease in survival was observed between patients with a score of −1 and those with a score of 0 (log-rank P=0.006; Supplementary Fig. 9B). Based on this finding, patients were classified into a simple-low (score <0) and simple-high (score ≥0) risk group. This simple 8-item risk scoring system successfully stratified the R0 gastrectomy group into two groups, with a significant difference in OS (log-rank P<0.001; Fig. 5B). Importantly, the simple 8-item risk scoring system not only demonstrated independent predictive performance (C-index, 0.74; Fig. 5C) but also showed high concordance with the original GBSA model’s risk stratification (accuracy=0.975, F1-score=0.975; Fig. 5C).
Fig. 5. Development and validation of the simple 8-item risk scoring system.
(A) Brief explanation of the simple 8-item risk scoring system. A point value of +1 was assigned for each adverse predictor (preoperative perforation, ASA status III or higher, preoperative obstruction) and −1 for each favorable predictor (upper body tumor, Borrmann type I or II, well or moderate differentiation, preoperative bleeding, and Krukenberg tumor). The final 8-item risk score is the sum of these points. For example, patient A, with only one adverse factor (preoperative obstruction), has a final score of +1. Patient B, with one favorable factor (Krukenberg tumor) and one adverse factor, has a final score of 0. Patient C, with two favorable factors, has a final score of −2. (B) Kaplan-Meier curve for OS between the simple-low group and simple-high group classified by the simple 8-item risk scoring system. p-values were calculated using the log-rank test. (C) Confusion matrix between the risk groups classified by the simple 8-item risk scoring system and the original GBSA model, with accuracy, F1-score, and the C-index of the simple 8-item risk scoring system presented.
ASA = American Society of Anesthesiologists; GBSA = gradient-boosting survival analysis; OS = overall survival; Pre op = preoperative.
DISCUSSION
Intraoperative management of peritoneal oligometastatic GC discovered during surgery remains a significant clinical challenge. Although conversion surgery following palliative chemotherapy is a favorable treatment paradigm for selected patients, a substantial proportion of patients are not candidates for this approach, leaving surgeons to make critical decisions regarding the extent of resection without objective guidance [1,2,3]. In addition, this approach is not always feasible in clinical settings, such as when systemic chemotherapy is not feasible or when urgent surgical intervention is required for complications such as bleeding, obstruction, or perforation. Accordingly, establishing a clear surgical strategy for this patient population, including patient selection, procedural approach, and the goal and extent of resection, remains a critical and urgent need.
To directly address this clinical dilemma, our recent large-scale multicenter study (PASS-META) provided compelling evidence regarding the oncologic value of gastrectomy in patients with GC and limited peritoneal metastasis (P1/P2) [6]. The current study extended this evidence from population-level associations to individual-level prognostic stratification. We successfully identified two distinct prognostic groups by applying a ML-based risk-stratification algorithm. Patients in the high-risk group had survival outcomes comparable to those of patients in the R2 gastrectomy or no gastrectomy groups. In contrast, patients in the low-risk group demonstrated significantly longer survival than those in the high-risk and control groups, with a 5-year survival rate of >30%. This outcome compares favorably with the survival rates reported in other major trials of stage IV GC, such as the DRAGON (~20%) and PHOENIX (~10%) studies [12,13]. These findings suggest that our stratification approach may help inform surgical decision-making by distinguishing patients with favorable prognoses after R0 gastrectomy from those whose outcomes are no better than those associated with noncurative approaches. However, as this was a retrospective, non-randomized comparison, the observed survival differences between low-risk patients and control groups may be confounded by indication bias and unmeasured variables and should not be interpreted as evidence of a causal treatment benefit. In practical terms, the decision-making process can be conceptualized as a 2-step approach: 1) assessing the technical feasibility of gastric resection and complete peritoneal metastasectomy, and 2) applying a simple score threshold to determine whether to proceed with upfront R0 gastrectomy or adopt a palliative strategy. To facilitate clinical application, this principle was further refined using our simple 8-item risk scoring system. A particularly notable finding was that patients with a score of 0, who lacked adverse features but did not possess a sufficient accumulation of favorable features, were categorized as high-risk. This finding highlights that the mere absence of adverse factors is insufficient to justify R0 gastrectomy. Instead, aggressive surgery should be reserved for carefully selected patients who exhibit distinct favorable prognostic indicators, such as an upper-body tumor, Borrmann type I or II tumor, or the presence of a Krukenberg tumor. This suggests that a “benefit-driven” approach, seeking positive reasons to operate, is superior to a “risk-avoidance” approach in this specific population.
The biological rationale for these favorable features is supported by the distinct oncological outcomes observed in our study. Patients in the low-risk group had significantly longer peritoneal PFS, and even after recurrence, they demonstrated longer post-progression survival compared with patients in the high-risk group. A possible explanation is that, in low-risk patients, the visible peritoneal seeding observed during surgery may not reflect a high burden of microscopic dissemination, resulting in a more favorable disease course after R0 gastrectomy. This hypothesis is supported by our findings regarding preoperative bleeding. As an alarming symptom, bleeding may prompt earlier diagnosis when the overall peritoneal disease burden remains limited, thereby enabling successful R0 gastrectomy and improving long-term outcomes. Furthermore, from a biological standpoint, several key features identified by our model appear to represent inherently less aggressive tumor behavior or limited peritoneal tumor burden. This interpretation is reinforced by our sub-analysis of the no-gastrectomy group, in which well- or moderately differentiated histology was associated with a significant survival advantage, and Borrmann type I or II tumors showed a strong trend toward improved survival (Supplementary Fig. 10). Interestingly, our model also identified upper-body tumor location as a favorable prognostic factor. Recent studies have reported that upper-body GCs have a significantly lower risk of peritoneal metastasis compared with distal cancers [14,15,16]. This inherent difference in metastatic pattern supports our findings. Therefore, upper-body tumor location in our model likely represents a subtype with limited peritoneal dissemination, which explains its association with a favorable prognosis. In addition, we found that the presence of a Krukenberg tumor was a favorable prognostic factor. This observation is consistent with several previous studies reporting that a significant survival benefit can be achieved through metastasectomy in highly selected patients with Krukenberg tumors, particularly when the disease is confined to the ovaries and a complete R0 gastrectomy is feasible [17,18].
In addition to identifying favorable characteristics, our model confirmed several well-established poor prognostic factors, further grounding its predictions in clinical practice. Preoperative perforation is a strong predictor of poor outcomes. This finding is clinically intuitive because perforation causes direct and widespread dissemination of tumor cells throughout the peritoneal cavity, likely resulting in a high microscopic disease burden that cannot be eliminated by R0 gastrectomy alone [19]. Consistent with its established role as a prognostic factor in primary GC surgery, our model confirmed that ASA physical status class III or higher was a powerful predictor of poor outcomes in patients with peritoneal oligometastatic disease [20,21]. A common concern regarding such high-risk features is that poor prognosis may be attributable to lower compliance with adjuvant chemotherapy owing to the patient's poor general condition. However, our analysis revealed no significant differences in the administration rate or total number of cycles of postoperative chemotherapy between the high- and low-risk groups. This is a critical finding, as it suggests that poor survival in the high-risk group is not merely a result of undertreatment but rather reflects the intrinsic biological aggressiveness of the tumor or the host-tumor interaction. A high ASA physical status classification in this context may represent not only comorbidities but also cancer-related systemic inflammation or cachexia driven by a high tumor burden, which creates a microenvironment favorable for recurrence despite standard adjuvant therapy. Preoperative obstruction was also identified as an adverse prognostic factor. Obstruction typically results from a locally advanced tumor in the lower body of the stomach. This is an important observation because, as supported by previous studies, noncardiac GCs have a significantly higher intrinsic risk of diffuse peritoneal dissemination [14,15,16]. Therefore, we hypothesized that peritoneal oligometastases in patients with obstructive tumors may not represent a truly localized disease state. In such cases, even a technically successful R0 gastrectomy of the visible nodules is likely to be futile because of the extensive underlying microscopic disease, which explains the poor prognosis associated with this feature in our model.
In this study, we performed a ML-based survival analysis using a component-wise gradient boosting survival model. Traditional survival analysis methods, such as the Cox proportional hazards model and Kaplan–Meier estimation, rely on a limited number of clinical variables and often fail to capture the complexity of patient heterogeneity [22,23,24]. These conventional approaches are particularly inadequate for identifying clinically meaningful subtypes of GC with peritoneal metastasis, in which survival outcomes are influenced by multiple interacting clinical features. Moreover, whereas the Cox model assumes proportional hazards over time and has difficulty accommodating nonlinear interactions among covariates, ML algorithms, including component-wise gradient boosting survival, provide a more flexible approach [25,26]. Indeed, a direct comparison between the GBSA and Cox models in our external validation cohort demonstrated that the GBSA model achieved a higher C-index (0.811 vs. 0.758), with superiority observed in 83% of bootstrapped resamples. By integrating perioperative and intraoperative clinical factors, this method enables dynamic risk assessment without relying on the proportional hazards assumption. Its ability to model complex nonlinear interactions among variables further enhances its predictive accuracy, making it a promising tool for refined prognostic stratification. To date, this study included the largest multicenter cohort of patients with GC and peritoneal metastasis. To ensure the generalizability of our survival model, we used patient data from four institutions for model training and data from two independent institutions for external validation. This approach allowed us to evaluate the robustness of the model in an independent patient population, confirming its prognostic reliability beyond that of the training dataset.
This study has certain limitations. First, as a multicenter retrospective analysis encompassing six institutions over seven years, it is subject to inherent biases, including the potential influence of unmeasured confounding variables. For example, differences in treatment protocols precluded the collection of detailed chemotherapy regimen data. Importantly, our model was prognostic rather than predictive. It stratifies patients according to the expected outcome after R0 gastrectomy but does not formally demonstrate a treatment-by-risk interaction. The observed survival differences between the low-risk patients and control groups (R2 gastrectomy or no gastrectomy) may reflect indication bias rather than a true differential treatment effect. Therefore, a large-scale prospective study is required to validate the clinical utility of our model and scoring system. Until such prospective validation is available, the model should be interpreted as a framework for prognostic risk communication rather than as a definitive tool for determining surgical candidacy, and clinical decisions regarding R0 gastrectomy should continue to rely on a comprehensive multidisciplinary assessment. Second, we used the Japanese Research Society for Gastric Cancer classification instead of the peritoneal cancer index score [9]. However, considering that the PASS-META study successfully validated the prognostic value of the P1/P2 classification in a large-scale population, we believe that our approach provides a practical surrogate for tumor burden in this setting [6]. Nevertheless, the P1/P2 classification does not capture the fine gradations of tumor burden within each category, and outcomes may differ even within P2 disease, depending on the extent of peritoneal involvement. Accordingly, patients with P2 amenable to R0 resection likely represent a subset with a lower tumor burden than those undergoing R2 resection or nonsurgical management. The favorable outcomes observed in the low-risk group should not be extrapolated to infer a survival benefit of gastrectomy across the broader P2 population. Third, significant selection bias was unavoidable. Because standardized treatment strategies for peritoneal metastatic GC have not yet been established, surgical decision-making across participating institutions was not uniform, and indications for staging laparoscopy, intraoperative management of peritoneal dissemination, and decisions to proceed with R0 resection varied according to institutional practice and surgeon judgment. As a result, the study cohort was limited to patients with peritoneal metastasis who were considered suitable candidates for curative-intent R0 gastrectomy, representing a highly selected and favorable subgroup within the broader population of patients with peritoneal metastasis. Therefore, the model may identify the characteristics of good surgical candidates as much as it identifies prognostic factors, and its findings may not be generalizable to patients whose tumors are deemed unresectable. Fourth, although this was a multicenter study, the entire cohort consisted of patients of a single ethnicity, and the small number of patients in the R0 gastrectomy group may have limited the generalizability and statistical power of the findings. The biology of GC differs across ethnic populations, and future validation of our model in larger, multiethnic international cohorts is warranted to confirm its broader applicability and external validity. Despite these limitations, the distinctive strength of our study lies in the unique cohort derived from the PASS-META study. Over a 7-year period across six major Korean centers, only 80 patients underwent R0 gastrectomy. This rarity likely reflects a clinical shift away from aggressive surgical approaches following the REGATTA trial [7]. Consequently, the difficulty of reproducing such a dataset in the current era underscores the significance and contribution of our analysis.
In conclusion, our study successfully developed and validated a ML model and a corresponding simple 8-item risk scoring system that effectively stratified patients with GC and peritoneal oligometastasis according to their prognosis after R0 gastrectomy. Given the retrospective observational design, the proposed framework should be regarded as a prognostic tool that may support, rather than determine, intraoperative decision-making; prospective validation is warranted before clinical implementation.
Footnotes
Funding: This research was supported by a grant from the Patient-Centered Clinical Research Coordinating Center (PACEN) funded by the Ministry of Health and Welfare, Republic of Korea (grant number: RS-2023-KH134639, RS-2026-25517509), and the Korea Health Industry Development Institute (KHIDI), Republic of Korea (grant number: RS-2025-02253005).
Conflict of Interest: No potential conflict of interest relevant to this article was reported.
- Conceptualization: L.S., J.Y.S., K.H.I.
- Data curation: S.H.J, S.J.H., K.S.E., L.I.S., L.H.H., J.O., J.M.R., H.H., J.Y.S., K.H.I.
- Formal analysis: L.S., K.H.I.
- Funding acquisition: L.S., H.H., K.H.I.
- Methodology: L.S.
- Project administration: K.H.I.
- Resources: S.H.J, S.J.H., K.S.E., L.I.S., L.H.H., J.O., J.M.R., H.H., J.Y.S., K.H.I.
- Software: L.S.
- Supervision: K.H.I.
- Validation: L.S., R.Y.H., K.H.I.
- Visualization: L.S.
- Writing - original draft: L.S.
- Writing - review & editing: L.S., R.Y.H., S.H.J, S.J.H., K.S.E., L.I.S., L.H.H., J.O., J.M.R., H.H., J.Y.S., K.H.I.
SUPPLEMENTARY MATERIALS
Baseline characteristics of patients stratified by treatment group and peritoneal metastasis stage
Comparison of baseline characteristics between training and validation cohorts in the R0 gastrectomy group
Schematic of peritoneal metastasis staging.
Survival analyses in the whole cohort (n=792).
Survival analyses of gastrectomy cohort in patients with P1/2/3 (n=177).
Survival analyses of gastrectomy cohort in patients with P1/2 (n=141).
Time-dependent AUC analysis of the GBSA model and Calibration.
Time-dependent AUC analysis of the GBSA model and Calibration.
Risk stratification using the ML model and its prognostic validation.
Comparison of PostOp systemic CTx compliance between risk groups.
Development and validation of the simple 8-item risk scoring system.
Survival analyses of no gastrectomy group (n=571).
References
- 1.Yang Z, Lu S, Shi M, Yuan H, Wang Z, Ni Z, et al. Oncological outcomes of conversion therapy in gastric cancer patients with peritoneal metastasis: a large-scale retrospective cohort study. Gastric Cancer. 2024;27:387–399. doi: 10.1007/s10120-023-01452-8. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 2.Chan KS, Xiao L, Oo AM. Outcomes of conversion surgery vs conventional systemic therapy in stage IV gastric cancer: a systematic review and meta-analysis. J Gastrointest Cancer. 2025;56:161. doi: 10.1007/s12029-025-01265-1. [DOI] [PubMed] [Google Scholar]
- 3.Yoshida K, Yasufuku I, Terashima M, Young Rha S, Moon Bae J, Li G, et al. International retrospective cohort study of conversion therapy for stage IV Gastric Cancer 1 (CONVO-GC-1) Ann Gastroenterol Surg. 2021;6:227–240. doi: 10.1002/ags3.12515. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 4.Kim TH, Uyama I, Rha SY, Bencivenga M, An J, Wyrwicz L, et al. Conversion therapy for stage IV gastric cancer: report from the expert consensus meeting at KINGCA WEEK 2024. J Gastric Cancer. 2025;25:133–152. doi: 10.5230/jgc.2025.25.e9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 5.Stephens MR, Lewis WG, White S, Blackshaw GRJC, Edwards P, Barry JD, et al. Prognostic significance of alarm symptoms in patients with gastric cancer. Br J Surg. 2005;92:840–846. doi: 10.1002/bjs.4984. [DOI] [PubMed] [Google Scholar]
- 6.Shin HJ, Song JH, Kim SE, Lee IS, Kim HI, Lee HH, et al. Clinical impact of gastrectomy in surgically proven stage IV gastric cancers: retrospective analysis from Korean multicenter dataset (PASS-META) Gastric Cancer. 2026;29:205–219. doi: 10.1007/s10120-025-01676-w. [DOI] [PubMed] [Google Scholar]
- 7.Fujitani K, Yang HK, Mizusawa J, Kim YW, Terashima M, Han SU, et al. Gastrectomy plus chemotherapy versus chemotherapy alone for advanced gastric cancer with a single non-curable factor (REGATTA): a phase 3, randomised controlled trial. Lancet Oncol. 2016;17:309–318. doi: 10.1016/S1470-2045(15)00553-7. [DOI] [PubMed] [Google Scholar]
- 8.Kim HI, Badgwell BD. Peritoneal oligometastasis in gastric cancer: diagnostic strategies, patient selection, and emerging therapeutic approaches. J Gastric Cancer. 2025;25:409–423. doi: 10.5230/jgc.2025.25.e36. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 9.Kajitani T. The general rules for the gastric cancer study in surgery and pathology. Part I. Clinical classification. Jpn J Surg. 1981;11:127–139. doi: 10.1007/BF02468883. [DOI] [PubMed] [Google Scholar]
- 10.Hothorn T, Bühlmann P, Dudoit S, Molinaro A, van der Laan MJ. Survival ensembles. Biostatistics. 2006;7:355–373. doi: 10.1093/biostatistics/kxj011. [DOI] [PubMed] [Google Scholar]
- 11.Moons KGM, Altman DG, Reitsma JB, Ioannidis JPA, Macaskill P, Steyerberg EW, et al. Transparent Reporting of a multivariable prediction model for Individual Prognosis or Diagnosis (TRIPOD): explanation and elaboration. Ann Intern Med. 2015;162:W1–W73. doi: 10.7326/M14-0698. [DOI] [PubMed] [Google Scholar]
- 12.Yan C, Yang Z, Shi Z, Lu S, Shi M, Nie M, et al. Intraperitoneal and intravenous paclitaxel plus S-1 versus intravenous paclitaxel plus S-1 in gastric cancer patients with peritoneal metastasis: Results from the multicenter, randomized, phase 3 DRAGON-01 trial. J Clin Oncol. 2025;43:327–327. [Google Scholar]
- 13.Ishigami H, Fujiwara Y, Fukushima R, Nashimoto A, Yabusaki H, Imano M, et al. Phase III trial comparing intraperitoneal and intravenous paclitaxel plus S-1 versus cisplatin plus S-1 in patients with gastric cancer with peritoneal metastasis: PHOENIX-GC trial. J Clin Oncol. 2018;36:1922–1929. doi: 10.1200/JCO.2018.77.8613. [DOI] [PubMed] [Google Scholar]
- 14.Wang L, Liang B, Jiang Y, Huang G, Tang A, Liu Z, et al. Subsite-specific metastatic organotropism and risk in gastric cancer: a population-based cohort study of the US SEER database and a Chinese single-institutional registry. Cancer Med. 2023;12:19595–19606. doi: 10.1002/cam4.6583. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 15.Rijken A, Lurvink RJ, Luyer MDP, Nieuwenhuijzen GAP, van Erning FN, van Sandick JW, et al. The burden of peritoneal metastases from gastric cancer: a systematic review on the incidence, risk factors and survival. J Clin Med. 2021;10:4882. doi: 10.3390/jcm10214882. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 16.Wei J, Yang Y, Du J, Zou Z, Zhang Y, Wu N, et al. Peritoneal metastasis in relation to outcome and therapeutic strategy in gastric cancer. Transl Cancer Res. 2017;6:149–156. [Google Scholar]
- 17.Cho JH, Lim JY, Choi AR, Choi SM, Kim JW, Choi SH, et al. Comparison of surgery plus chemotherapy and palliative chemotherapy alone for advanced gastric cancer with Krukenberg tumor. Cancer Res Treat. 2015;47:697–705. doi: 10.4143/crt.2013.175. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 18.Ma F, Li Y, Li W, Kang W, Liu H, Ma S, et al. Metastasectomy improves the survival of gastric cancer patients with Krukenberg tumors: A retrospective analysis of 182 patients. Cancer Manag Res. 2019;11:10573–10580. doi: 10.2147/CMAR.S227684. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 19.Lee MH, Kim YM, Oh MK, Kim HI, Kim MG. Risk factor analysis and surgical outcomes in emergency gastrectomy for gastric perforation: a single-center retrospective cohort study. Ann Surg Treat Res. 2025;109:185–193. doi: 10.4174/astr.2025.109.3.185. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 20.Nishibeppu K, Sakuramoto S, Matsui K, Ebara G, Fujita S, Fujihata S, et al. Dismal prognosis of elderly gastric cancer patients who underwent gastrectomy with American Society of Anesthesiologists (ASA) 3. Langenbecks Arch Surg. 2022;407:3413–3421. doi: 10.1007/s00423-022-02672-9. [DOI] [PubMed] [Google Scholar]
- 21.Rosa F, Tortorelli AP, Quero G, Galiandro F, Fiorillo C, Sollazzi L, et al. The impact of preoperative ASA-physical status on postoperative complications and long-term survival outcomes in gastric cancer patients. Eur Rev Med Pharmacol Sci. 2019;23:7383–7390. doi: 10.26355/eurrev_201909_18846. [DOI] [PubMed] [Google Scholar]
- 22.Moncada-Torres A, van Maaren MC, Hendriks MP, Siesling S, Geleijnse G. Explainable machine learning can outperform Cox regression predictions and provide insights in breast cancer survival. Sci Rep. 2021;11:6968. doi: 10.1038/s41598-021-86327-7. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 23.Huang Y, Li J, Li M, Aparasu RR. Application of machine learning in predicting survival outcomes involving real-world data: a scoping review. BMC Med Res Methodol. 2023;23:268. doi: 10.1186/s12874-023-02078-1. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 24.Salerno S, Li Y. High-dimensional survival analysis: methods and applications. Annu Rev Stat Appl. 2023;10:25–49. doi: 10.1146/annurev-statistics-032921-022127. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 25.De Bin R, Stikbakke VG. A boosting first-hitting-time model for survival analysis in high-dimensional settings. Lifetime Data Anal. 2023;29:420–440. doi: 10.1007/s10985-022-09553-9. [DOI] [PMC free article] [PubMed] [Google Scholar]
- 26.Oh N, Choi GS, Lee WY. ChatGPT goes to the operating room: evaluating GPT-4 performance and its potential in surgical education and training in the era of large language models. Ann Surg Treat Res. 2023;104:269–273. doi: 10.4174/astr.2023.104.5.269. [DOI] [PMC free article] [PubMed] [Google Scholar]
Associated Data
This section collects any data citations, data availability statements, or supplementary materials included in this article.
Supplementary Materials
Baseline characteristics of patients stratified by treatment group and peritoneal metastasis stage
Comparison of baseline characteristics between training and validation cohorts in the R0 gastrectomy group
Schematic of peritoneal metastasis staging.
Survival analyses in the whole cohort (n=792).
Survival analyses of gastrectomy cohort in patients with P1/2/3 (n=177).
Survival analyses of gastrectomy cohort in patients with P1/2 (n=141).
Time-dependent AUC analysis of the GBSA model and Calibration.
Time-dependent AUC analysis of the GBSA model and Calibration.
Risk stratification using the ML model and its prognostic validation.
Comparison of PostOp systemic CTx compliance between risk groups.
Development and validation of the simple 8-item risk scoring system.
Survival analyses of no gastrectomy group (n=571).