Machine-learning-based prediction model for Clavien-Dindo grade ≥ II complications after neoadjuvant therapy and laparoscopic gastrectomy in gastric cancer

Abstract

BACKGROUND

Neoadjuvant therapy prior to surgery plays a critical role in improving the prognosis of patients with unresectable or locally advanced gastric cancer (GC). Postoperative complications, particularly those classified as Clavien-Dindo grade ≥ II, remain a major concern for surgeons. In recent years machine learning (ML) has emerged as a prominent approach for disease diagnosis and prediction. However, studies on both postoperative complications and ML in patients with GC receiving neoadjuvant therapy remain limited.

AIM

To develop an ML model to predict Clavien-Dindo grade ≥ II complications in patients with GC after neoadjuvant therapy and laparoscopic gastrectomy.

METHODS

Clinical data were collected from 455 patients with GC who underwent neoadjuvant therapy followed by laparoscopic gastrectomy at Peking Union Medical College Hospital (2014-2024). Potential predictors were identified through univariate analysis and least absolute shrinkage and selection operator regression. Six ML algorithms including XGBoost, random forest, neural network ensemble (NNE), logistic regression, GLMnet, and decision tree were trained and optimized using nested cross-validation. Model performance was evaluated using the area under the receiver operating characteristic curve, decision curve analysis, and calibration curves.

RESULTS

A total of 455 patients were included of whom 69 (15.16%) developed Clavien-Dindo grade ≥ II complications. The predictive model was constructed using seven variables, including smoking status, Nutritional Risk Screening-2002 score, American Society of Anesthesiologists classification, neoadjuvant therapy, surgical approach, operating time, and intraoperative blood loss. Among the six models the NNE model outperformed the others, achieving the highest area under the receiver operating characteristic curve (0.789, 0.739-0.840) and demonstrating superior discrimination, clinical utility, and calibration.

CONCLUSION

The NNE-based prediction model effectively identified patients with GC at high risk of Clavien-Dindo grade ≥ II complications after neoadjuvant therapy and laparoscopic gastrectomy.

Key Words: Gastric cancer; Machine learning; Postoperative complications; Risk prediction; Neoadjuvant therapy

Core Tip: Addressing data scarcity in gastric cancer neoadjuvant therapy, this study used a large patient cohort (n = 455) to pioneer a machine learning model for predicting Clavien-Dindo grade ≥ II complications post-neoadjuvant therapy and laparoscopic gastrectomy. Key predictors included smoking status, Nutritional Risk Screening-2002 score, American Society of Anesthesiologists classification, neoadjuvant therapy, surgical approach, operating time, and intraoperative blood loss. The neural network ensemble model demonstrated superior performance with optimal discrimination, calibration, and clinical utility, potentially offering a tool for perioperative risk stratification and management optimization.

INTRODUCTION

Gastric cancer (GC) is a common malignant tumor of the digestive system, ranking fifth in incidence and third in mortality among all cancers worldwide[1]. Its development is associated with multiple risk factors, including family history, smoking, alcohol consumption, dietary habits, Helicobacter pylori infection, and Epstein-Barr virus infection[2,3]. Notably, > 95% of GCs are adenocarcinomas[4]. The diagnosis of GC mainly relies on clinical symptoms, enhanced CT, positron emission tomography/CT, tumor markers, and gastroscopic biopsy[3]. Despite significant advancements in diagnostic techniques in recent years, most patients are diagnosed at an advanced stage due to the late onset of symptoms[5,6]. Currently, surgery remains the cornerstone of treatment for advanced GC[3]. However, with the evolution of multidisciplinary treatment approaches, neoadjuvant therapy has emerged as an increasingly important component of comprehensive management.

Neoadjuvant therapy, administered prior to surgery, encompasses chemotherapy, radiotherapy, targeted therapy, and immunotherapy[7,8]. Leading clinical guidelines from Japan, United States, Europe, and China recommend preoperative neoadjuvant therapy for patients with locally advanced or potentially resectable GC to improve prognosis[6,8-10]. Recently, research on novel neoadjuvant therapy regimens has become a key focus in GC treatment[11,12].

Despite significant advances in systemic treatment, postoperative complications remain a major challenge in GC. Based on the Clavien-Dindo classification system, 30-day postoperative complications are categorized into five grades: I (minor complications requiring no treatment); II (requiring pharmacological intervention); III (needing surgical/endoscopic/radiological intervention); IV (life-threatening complications requiring intensive care); and V (death). Grade ≥ II complications adversely affect patient recovery, prolong hospitalization time, and escalate healthcare costs and readmission rates. These impacts collectively impose significant burdens on patients and healthcare systems, emphasizing the necessity of effective complication prediction and prevention strategies to optimize outcomes and resource allocation.

Machine learning (ML), a rapidly advancing branch of artificial intelligence, has been increasingly applied in medical research, particularly in oncology for disease prediction and risk stratification[13,14]. While several ML-based studies have investigated postoperative complications in patients with GC[15,16], most have excluded those receiving neoadjuvant therapy, likely due to limited sample sizes in this specific population.

This study aimed to bridge this critical research gap by identifying influential preoperative and intraoperative factors associated with Clavien-Dindo grade ≥ II complications in patients with neoadjuvant-treated GC and developing an ML-based predictive model. The proposed model has significant clinical implications as it can stratify the risk of perioperative complications in patients with GC undergoing neoadjuvant therapy. It assists clinicians in optimizing treatment strategies to achieve precision medicine, thereby preventing complications and improving patient outcomes.

MATERIALS AND METHODS

Patient selection

This study retrospectively analyzed patients who underwent laparoscopic radical gastrectomy following neoadjuvant therapy at Peking Union Medical College Hospital from January 2014 to December 2024. Clinical data were systematically collected for all eligible cases. Inclusion criteria included: (1) Completion of preoperative neoadjuvant therapy; (2) Radical gastrectomy (D2 lymphadenectomy with R0 resection); (3) Histopathological confirmation of gastric adenocarcinoma; and (4) Procedures performed by experienced surgeons (minimum 50 prior cases of the same procedure). Exclusion criteria included: (1) Age < 18 or > 85 years; (2) Intraoperative findings of peritoneal metastasis, adjacent organ invasion, or distant metastasis; (3) Other malignancies; (4) Severe comorbidities (e.g., cardiovascular or cerebrovascular diseases); (5) Non-standardized chemotherapy regimen (e.g., < 2 treatment cycles); and (6) Loss to follow-up. The primary outcome was the occurrence of Clavien-Dindo grade ≥ II complications.

Data collection

Comprehensive patient data were systematically collected, encompassing demographic characteristics, preoperative clinical data, intraoperative data, and postoperative complications. Following adjustment for potential confounders, 34 clinically relevant factors were analyzed, including age, sex, preoperative hospital stay, smoking, alcohol consumption, hypertension, diabetes, gastroesophageal reflux disease, gastric outlet obstruction, psychological disorders, history of abdominal surgery, thyroid function, Helicobacter pylori infection, combination of other disease, family history of tumors, neoadjuvant therapy, Nutritional Risk Screening-2002 (NRS-2002) score, tumor location, clinical stage at initial diagnosis, American Society of Anesthesiologists (ASA) status, cardiac function, body mass index, preoperative WBC count, hemoglobin, glucose, albumin, albumin-to-globulin ratio, surgical approach, operating time, blood loss, intraoperative endoscopy use, implantation of nutritional tube, blood transfusion, and postoperative complication occurrence and classification.

Data preprocessing and variable selection

In this study variables with missing values exceeding 30% were excluded from analysis. For variables with missing values not exceeding 30%, we used multiple imputation for data completion. Specifically, we utilized the bootstrap method for resampling and applied the predictive mean matching algorithm to generate imputed values. All analytical variables were simultaneously incorporated into the imputation model to maintain intervariable relationships. Finally, we generated and pooled results from five imputed datasets to obtain the final estimates.

Statistical analysis

Univariate analyses were first performed with categorical variables analyzed using χ² or Fisher’s exact tests, normally distributed continuous variables compared using t tests (reported as mean ± SD), and non-normally distributed variables assessed using Wilcoxon rank-sum tests (reported as median; interquartile range). Variables with P < 0.05 were considered statistically significant. Subsequently, least absolute shrinkage and selection operator (LASSO) regression with 10-fold cross-validation was used to identify predictive features in which the optimal regularization parameter (λ) was determined through minimum criteria and variables with non-zero coefficients were retained. Finally, the intersection of significant variables from both univariate analysis and LASSO regression was used to construct the ML prediction models.

Model development and evaluation

Within the mlr3 framework utilizing nested cross-validation (fivefold inner loop and fivefold outer loop), we implemented six ML algorithms, XGBoost, random forest, neural network ensemble (NNE), logistic regression, GLMnet, and decision tree, to develop the predictive models. To address potential model bias arising from class imbalance, the synthetic minority oversampling technique was applied to preprocess the data within the inner loop of the cross-validation framework. Model performance was comprehensively assessed using multiple metrics including the area under the receiver operating characteristic curve (ROC), accuracy, specificity, and recall. Additional validation was performed through analysis of ROC curves, decision curve analysis (DCA), and calibration plots.

Model interpretation and feature importance

SHapley Additive exPlanations (SHAP) analysis was used to quantify and rank the relative contributions of predictive features in the optimal ML model, thereby facilitating identification of key risk factors for Clavien-Dindo grade ≥ II complications following neoadjuvant therapy in patients with GC. All statistical analyses were performed using R software (version 4.4.3). The workflow of this study is illustrated in Figure 1.

Figure 1 Flow chat for study design. LASSO: Least absolute shrinkage and selection operator; NNE: Neural network ensemble; ROC: Receiver operating characteristic; SHAP: SHapley Additive exPlanations.

RESULTS

Patient characteristics

This study included 455 patients with GC who received neoadjuvant therapy prior to radical gastrectomy, comprising 342 males (75.2%) and 113 females (24.8%) with a median age of 61 (52-68) years. The baseline characteristics, including demographic data, preoperative clinical data, surgical details, and postoperative complications are presented in Table 1. Clavien-Dindo grade ≥ II complications occurred in 69 patients (15.2%). The spectrum of complications included: Anastomotic leaks (n = 16, 3.5%), pulmonary infection (n = 16, 3.5%), hemorrhage (n = 12, 2.6%), gastroparesis (n = 12, 2.6%), anastomotic stricture (n = 4, 0.9%), intestinal obstruction (n = 3, 0.7%), venous thrombosis, incision infection, and lymphatic leaks (n = 2 each, 0.4%), pancreatic fistulas, bloodstream infections, myocardial infarction, postoperative adrenal crisis, atrial fibrillation, urinary tract infection, and pulmonary embolism (n = 1 each, 0.2%).

Table 1 Baseline characteristics and univariate analysis of patients (abridged version).

	All (N = 455)	No (n = 386)	Yes (n = 69)	Univariate analysis		N
				OR	P value
Age, year	61.0 (52.0-68.0)	61.0 (52.2-67.0)	62.0 (51.0-68.0)	1.0 (0.98-1.02)	0.809	455
Sex					0.045a	455
Female	113 (24.8)	103 (26.7)	10 (14.5)	Reference
Male	342 (75.2)	283 (73.3)	59 (85.5)	2.12 (1.09-4.57)
Smoking					< 0.001a	455
No	230 (50.5)	209 (54.1)	21 (30.4)	Reference
Yes	225 (49.5)	177 (45.9)	48 (69.6)	2.68 (1.56-4.75)
ASA status					< 0.001a	455
L1	54 (11.9)	48 (12.4)	6 (8.70)	Reference
L2	314 (69.0)	279 (72.3)	35 (50.7)	0.98 (0.42-2.75)
L3	87 (19.1)	59 (15.3)	28 (40.6)	3.70 (1.49-10.7)
Neoadjuvant therapy					0.008a	455
Chemotherapy	344 (75.6)	300 (77.7)	44 (63.8)	Reference
Radiochemotherapy	60 (13.2)	43 (11.1)	17 (24.6)	2.70 (1.38-5.10)
Combined immuno/targeted therapy	51 (11.2)	43 (11.1)	8 (11.6)	1.28 (0.53-2.80)
NRS-2002 ≥ 3					< 0.001a	455
No	294 (64.6)	265 (68.7)	29 (42.0)	Reference
Yes	161 (35.4)	121 (31.3)	40 (58.0)	3.01 (1.78-5.13)
Surgical approach					< 0.001a	455
Distal-Billroth 1	155 (34.1)	146 (37.8)	9 (13.0)	Reference
Distal-Billroth 2	6 (1.32)	4 (1.04)	2 (2.90)	8.11 (0.9-50.9)
Distal-Roux-en-Y	68 (14.9)	50 (13.0)	18 (26.1)	5.74 (2.46-14.3)
Proximal	8 (1.76)	7 (1.81)	1 (1.45)	2.54 (0.09-17.4)
Total-Roux-en-Y	218 (47.9)	179 (46.4)	39 (56.5)	3.48 (1.70-7.93)
Implantation of nutrition tubes					0.025a	455
No	237 (52.1)	192 (49.7)	45 (65.2)	Reference
Yes	218 (47.9)	194 (50.3)	24 (34.8)	0.53 (0.31-0.90)
Operating time > 4 h					< 0.001a	455
No	254 (55.8)	231 (59.8)	23 (33.3)	Reference
Yes	201 (44.2)	155 (40.2)	46 (66.7)	2.96 (1.74-5.17)
Blood loss > 100 mL					< 0.001a	455
No	321 (70.5)	289 (74.9)	32 (46.4)	Reference
Yes	134 (29.5)	97 (25.1)	37 (53.6)	3.43 (2.03-5.85)

^aP < 0.05.

Data are presented as n (%). The table above lists the patient demographics and all variables that were significant in the univariate analysis. The complete baseline characteristics are available in Supplementary Table 5. OR: Odds ratio; ASA: American Society of Anesthesiologists; NRS-2002: Nutritional Risk Screening-2002.

Data preprocessing and feature selection

Univariate analysis demonstrated significant differences between patients with Clavien-Dindo grade ≥ II complications and those without in the following variables: Sex (P = 0.045); smoking status (P < 0.001); ASA status (P < 0.001); neoadjuvant therapy (P = 0.008); NRS-2002 score ≥ 3 (P < 0.001); surgical approach (P < 0.001); nutritional tube placement (P = 0.025); operating time > 4 h (P < 0.001); and blood loss > 100 mL (P < 0.001).

Figure 2 presents the LASSO regression results at the optimal λ value, identifying several clinically significant predictors of Clavien-Dindo grade ≥ II complications: Higher ASA status (grade 3, coefficient = 0.814); prolonged operating time (> 4 h, coefficient = 0.640); increased blood loss (> 100 mL, coefficient = 0.597); elevated nutritional risk (NRS-2002 score ≥ 3, coefficient = 0.584); current smoking status (coefficient = 0.560); combined neoadjuvant chemotherapy with target/immunotherapy (coefficient = 0.463); Roux-en-Y reconstruction following distal gastrectomy (coefficient = 0.293); tumor location (proximal, coefficient = 0.122); and longer preoperative hospitalization (coefficient = 0.021). Normal albumin-to-globulin ratio demonstrated a protective effect (coefficient = -0.447).

Figure 2 Least absolute shrinkage and selection operator regression results. A: Least absolute shrinkage and selection operator regression cross-validation curve; B: Key variables for least absolute shrinkage and selection operator regression screening. LASSO: Least absolute shrinkage and selection operator; ASA: American Society of Anesthesiologists; NRS: Nutritional Risk Screening; A/G: Albumin-to-globulin ratio.

Based on the combined results of univariate and LASSO analyses, the following variables were selected for ML model construction: Smoking; NRS-2002 score ≥ 3; ASA status; neoadjuvant therapy; surgical strategy; operating time > 4 h; and blood loss > 100 mL. No significant multicollinearity was detected among these predictors. The correlation heatmap illustrating the relationships between these variables is provided in Supplementary Figure 1. The sample size in this study was determined in accordance with established guidelines for predictive model research. Based on the events per variable criterion, 455 patients were included, among whom 69 experienced postoperative complications (positive events). Following feature selection, the model incorporated seven predictor variables. The resulting events per variable ratio was approximately 9.86:1, which was close to the commonly recommended threshold of 10:1, indicating that the current sample size is sufficient to ensure the stability and reliability of the model construction.

Development and validation of ML models

Using seven clinically significant predictors (smoking status, NRS-2002 score ≥ 3, ASA status, neoadjuvant therapy, surgical approach, operating time > 4 h, and blood loss > 100 mL), we developed and validated six ML models (XGBoost, random forest, NNE, logistic regression, GLMnet, and decision tree). NNE demonstrated superior performance, achieving the highest area under the curve (AUC = 0.789) and a high recall value (0.731) (Table 2). Logistic regression also showed competitive predictive accuracy with an AUC of 0.786.

Table 2 Performance metrics of the models.

Model	AUC	Accuracy	Recall	Sensitivity	Specificity
NNE	0.789 (0.739-0.840)	0.732 (0.657-0.807)	0.731 (0.610-0.851)	0.731 (0.610-0.851)	0.733 (0.625-0.841)
Logistic regression	0.786 (0.606-0.966)	0.752 (0.697-0.807)	0.763 (0.525-0.824)	0.763 (0.525-0.824)	0.756 (0.704-0.807)
GLMnet	0.776 (0.668-0.884	0.67 (0.555-0.786)	0.739 (0.570-0.908)	0.739 (0.570-0.908)	0.656 (0.508-0.803)
Random forest	0.754 (0.648-0.861)	0.666 (0.641-0.691)	0.735 (0.505-0.965)	0.735 (0.505-0.965)	0.659 (0.598-0.720)
XGBoost	0.703 (0.571-0.835)	0.679 (0.567-0.791)	0.606 (0.400-0.813)	0.606 (0.400-0.813)	0.693 (0.577-0.810)
Decision tree	0.602 (0.544-0.660)	0.756 (0.670-0.842)	0.371 (0.259-0.482)	0.371 (0.259-0.482)	0.827 (0.725-0.928)

NNE: Neural network ensemble; AUC: Area under the curve.

ROC curve analysis confirmed the optimal discriminative ability of NNE (AUC = 0.789), followed by logistic regression (AUC = 0.786) and GLMnet (AUC = 0.776) (Figure 3). DCA revealed that NNE provided the greatest clinical net benefit across threshold probabilities, outperforming logistic regression and GLMnet (Figure 4). Calibration plots indicated that the predictions of NNE best approximated the ideal diagonal, demonstrating excellent calibration accuracy (Figure 5). In summary, NNE emerged as the most robust predictive model for Clavien-Dindo grade ≥ II complications following neoadjuvant therapy in GC patients, combining high discriminative power (AUC), clinical utility (DCA), and calibration accuracy.

Figure 3 Receiver operating characteristic curves of the machine learning prediction models. ROC: Receiver operating characteristic; AUC: Area under the curve.

Figure 4  Decision curve analysis curves of the machine learning prediction models.

Figure 5  Calibration curves of the machine learning prediction models.

SHAP-based feature importance analysis

SHAP analysis of the optimal NNE model (Figure 6) revealed the following ranked order of predictive features for Clavien-Dindo grade ≥ II complications: (1) NRS-2002 score > 3 (risk factor); (2) Operating time ≤ 4 h (protective factor); (3) Smoking (risk factor); (4) ASA-L2 (protective factor); (5) Distal gastrectomy with Billroth I anastomosis (protective factor); (6) Blood loss ≤ 100 mL (protective factor); and (7) Neoadjuvant chemotherapy (protective factor).

Figure 6 SHapley Additive exPlanations feature importance in neural network ensemble. NNE: Neural network ensemble; SHAP: SHapley Additive exPlanations; NRS: Nutritional Risk Screening; ASA: American Society of Anesthesiologists.

Subgroup analysis

Based on the aforementioned SHAP analysis results, we performed subgroup analysis focusing on the type of neoadjuvant therapy. The cohort was stratified into three subgroups: Neoadjuvant radiotherapy alone, neoadjuvant chemoradiotherapy, and neoadjuvant chemotherapy combined with immunotherapy or targeted therapy. Propensity score matching was applied to compare these subgroups in a pairwise manner. The baseline characteristics after matching are presented in Supplementary Tables 1-3. χ² tests conducted on the matched data revealed no significant differences in the incidence of postoperative complications among the different neoadjuvant therapy regimens (Supplementary Table 4).

DISCUSSION

Neoadjuvant therapy has become an established treatment for locally advanced GC, demonstrating efficacy in tumor downstaging and improved prognosis[17,18]. However, research on postoperative complications following neoadjuvant-treated resection remain limited. Given ML proven utility in predicting oncological outcomes[19,20], this study leveraged six ML algorithms within the mlr3 framework utilizing nested cross-validation for enhanced stability, rigorous leakage prevention, and robust evaluation in limited sample sizes[21] to analyze Clavien-Dindo grade ≥ II complications based on preoperative/intraoperative variables. The developed predictive models identified key risk factors, offering clinical utility for preoperative risk stratification, intraoperative decision-making, and early postoperative management to mitigate complications and optimize recovery.

This study identified smoking status, NRS-2002 score ≥ 3, ASA status, neoadjuvant therapy, surgical approach, operating time > 4 h, and blood loss > 100 mL as significant predictors of Clavien-Dindo grade ≥ II complications following neoadjuvant therapy for GC. Among the six ML models, the NNE model demonstrated superior predictive performance, achieving the highest AUC (0.789) compared with conventional logistic regression (AUC = 0.786). The NNE model also exhibited optimal clinical utility on DCA and maintained excellent calibration accuracy. SHAP value analysis further revealed the relative importance of predictive features in the NNE model ranked in descending order: NRS-2002 score; operating time; smoking status; ASA status; surgical approach; blood loss; and neoadjuvant therapy. It is easy to find that surgical factors represent the most significant determinants of Clavien-Dindo grade ≥ II complications in patients with GC patients following neoadjuvant therapy.

Our findings demonstrate that both distal and total gastrectomy with Roux-en-Y reconstruction are associated with significantly higher rates of Clavien-Dindo grade ≥ II complications compared with distal gastrectomy with Billroth I reconstruction, a result that aligns with the 2024 multicenter study (n = 2508) evaluating complication risk factors in GC surgery[22]. This association may be explained by the increased technical complexity and greater tissue trauma associated with this anastomotic technique. The study further demonstrated that intraoperative blood loss > 100 mL serves as a risk factor, corroborating results reported by Yu et al[22]. The underlying pathophysiology likely involves impaired gastrointestinal perfusion during significant hemorrhage, potentially increasing the risk of anastomotic leakage and related complications[23].

Additionally, operating time > 4 h was significantly associated with postoperative complications in this study. These findings align with previous research demonstrating that surgical procedures > 4 h correlated with higher complication rates and infection rates following radical gastrectomy[24,25]. The cumulative evidence suggests that surgical management of neoadjuvant-treated GC should incorporate three key strategies: Using a simplified surgical approach when feasible; optimizing surgical time without compromising radical resection standards; and implementing stringent blood conservation measures including judicious use of hemostatics and transfusion. This comprehensive approach may significantly reduce postoperative complication risks while ensuring optimal surgical outcomes.

The NRS-2002 score is a commonly used clinical tool for nutritional screening in patients with GC with a score ≥ 3 indicating nutritional risk[26]. This paper demonstrated that an NRS-2002 score ≥ 3 is a risk factor for postoperative complications in patients with GC receiving neoadjuvant therapy. Previous research has shown that an NRS-2002 score ≥ 5 is an independent risk factor for postoperative infections in patients with GC[27] while malnutrition is associated with increased complications and poor prognosis following various oncological operations[28,29]. These findings align with the results of the present study, underscoring the necessity of perioperative nutritional assessment. For patients with GC undergoing neoadjuvant therapy who are at nutritional risk, nutritional support therapy should be implemented to prevent postoperative complications and improve clinical outcomes.

ASA status serves as a critical preoperative risk assessment tool with existing literature demonstrating a significant correlation between higher ASA status (III/IV) and increased postoperative complication rates in GC surgery[15,30], a finding consistent with our study results. While this association may be partially mediated through elevated infection risks secondary to impaired physiological reserve, the precise mechanistic pathways warrant further investigation through prospective, pathophysiology-focused studies.

Substantial evidence has established smoking as a significant risk factor for GC development[31,32] with studies demonstrating a dose-response relationship in which prolonged smoking time increases risk while extended smoking cessation reduces susceptibility[33]. The current study further identified smoking as a risk factor for postoperative complications in patients with GC receiving neoadjuvant therapy. These findings suggest that smoking not only contributes to gastric carcinogenesis but also adversely impacts surgical outcomes although the underlying mechanisms warrant further investigation. Importantly, these results highlight the critical need for comprehensive smoking cessation counseling prior to surgery in patients who smoke and undergo neoadjuvant therapy as this intervention may significantly reduce the risk of postoperative complications.

In addition, the results of this study suggest that patients receiving neoadjuvant therapy combined with immunotherapy or targeted therapy may be at a higher risk of postoperative complications, representing a novel finding. To clarify whether the observed association between different neoadjuvant therapy and postoperative complications was attributable to the treatments themselves or to patient selection bias, we conducted a subgroup analysis based on the specific neoadjuvant therapy received. The results demonstrated that after propensity score matching there were no significant differences in postoperative complication rates among the various neoadjuvant therapy subgroups, a finding consistent with previous studies[34-36]. This suggests that the impact of neoadjuvant therapy identified by our model is primarily driven by patient selection, meaning that patients receiving neoadjuvant therapy combined with immunotherapy or targeted therapy inherently possessed higher baseline risks rather than being a direct effect of the therapies themselves. This conclusion is supported by our SHAP analysis, which indicated that the type of neoadjuvant therapy had a low overall importance ranking in the model. Although the subgroup analysis suggested that immunotherapy and targeted therapy are not independent risk factors for postoperative complications, retaining this variable in the model contributed to enhancing its overall discriminative ability. Therefore, we opted to include this feature in the final model.

The findings of this study demonstrated that in terms of overall discriminative ability (AUC) both our proposed NNE model (AUC = 0.789) and logistic regression model (AUC = 0.786) achieved superior performance, significantly outperforming the other comparative models. However, the NNE model exhibited a narrower AUC confidence interval (0.739-0.840) vs the logistic regression model (0.606-0.966), indicating stronger and more stable overall classification and discriminative capability. The NNE model (0.731) performed comparably to logistic regression (0.763), GLMnet (0.739), and random forest (0.735) but significantly better than XGBoost (0.606) and decision tree (0.371).

Concurrently, the NNE model again showed a narrower confidence interval (0.739-0.840), suggesting its excellence and stability in positive case identification. Importantly, this comparative analysis extended beyond traditional performance metrics. In terms of clinical utility, DCA revealed that the NNE model provided the highest net clinical benefit across the vast majority of threshold probability ranges. This implies that using the NNE model for clinical decision-making predictions can yield better expected outcomes for patients compared with other models, underscoring its high practical value.

Regarding predictive reliability, the calibration curve indicated that the predicted probabilities of the NNE model align most closely with the actual observed probabilities with the curve tightly following the ideal reference line, demonstrating the best calibration. This suggests that the predictions made by the NNE model are not only accurate (good discriminative performance) but also highly reliable in terms of the predicted probability values themselves. In summary, through a multifaceted and detailed comparative analysis, the results confirmed that our proposed NNE model not only excels in discriminative performance but also demonstrated significant advantages in clinical utility and predictive reliability, thereby robustly validating its advanced nature, effectiveness, and practical value.

To enhance the practicality of the NNE model and facilitate future validation, we have shared the model data via a stable cloud repository (https://drive.google.com/file/d/1pDo_XeJv1HsZxDzNGaBsJbPhzye0ehjN/view?usp=sharing), providing rapid and real-time clinical reference for surgeons. A screenshot of the application interface is provided in Supplementary Figure 2. Based on the SHAP analysis and NNE predictive model, surgeons can estimate the probability of Clavien-Dindo grade ≥ II complications occurring in patients during the preoperative and early postoperative periods, enabling individualized risk assessment and precision medicine. For instance, according to the predictive model outcomes, surgeons may optimize perioperative clinical decision-making, such as implementing preoperative nutritional support, smoking cessation education, or selecting superior surgical approaches, thereby potentially preventing complications and improving patient prognosis.

This study had several limitations that should be acknowledged. The retrospective design and reliance on a single high volume center dataset may impact the generalization of our findings. Additionally, continuous variables such as operative time and blood loss were subject to measurement inaccuracies inherent in retrospective data collection, which may have resulted in loss of granularity. While these constraints are currently unavoidable due to limited availability of clinical data for patients with neoadjuvant-treated GC, we are implementing two key strategies to address these limitations: (1) Initiating prospective multicenter collaborative studies to enhance sample diversity and develop more robust predictive models; and (2) Sharing the model data via an open-access repository (https://drive.google.com/file/d/1pDo_XeJv1HsZxDzNGaBsJbPhzye0ehjN/view?usp=sharing) to facilitate external validation and provide clinical reference for surgical teams worldwide. This dual approach will significantly advance the translation of our predictive models into clinical practice while ensuring continuous model optimization through global data integration. Subsequent studies will systematically compare the diagnostic efficiency of this complication prediction model with conventional clinical risk scores such as Physiological and Operative Severity Score for the enUmeration of Mortality and morbidity and American College of Surgeons National Surgical Quality Improvement Project, representing a major focus of our future research agenda.

CONCLUSION

ML demonstrated significant potential for predicting Clavien-Dindo grade ≥ II complications in patients with GC receiving neoadjuvant therapy and laparoscopic radical gastrectomy. This study identified seven key predictive factors for Clavien-Dindo grade ≥ II complications: Smoking status; NRS-2002 score ≥ 3; ASA status; neoadjuvant therapy; surgical approach; operating time > 4 h; and blood loss > 100 mL. Among the predictive models developed using these factors, the NNE model demonstrated superior performance in predicting Clavien-Dindo grade ≥ II complications, achieving an AUC of 0.789 along with better discrimination, clinical utility, and calibration. With the implementation of the NNE prediction model web tool, it may facilitate the development of personalized precision medicine strategies to mitigate complications and optimize postoperative recovery in patients with GC.