Introduction

Gastric cancer is the third leading cause of cancer-related death worldwide [1]. Nevertheless, gastrectomy provides a considerable benefit on the survival of patients with gastric cancer and plays an important role in treatment. Gastric cancer incidence is high in East Asia; annually, around 52,000 patients undergo gastrectomy for gastric cancer treatment in Japan according to the Japanese National Clinical Database (NCD) [2, 3]. Abundant gastrectomies contribute to standardizing the surgical technique and perioperative management, leading to a low mortality of 1.1–2.3%. However, the morbidity rate remains high, accounting for 14.2% in distal gastrectomy and 21.5% in total gastrectomy [2, 3]. Hence, the morbidity rate should be reduced. Among the postgastrectomy complications, anastomotic leakage, intra-abdominal abscess, and pancreatic fistula are defined as postoperative intra-abdominal infectious complications (PIICs), and the incidences of PIIC are reportedly 3.3–10.6% [4, 5]. PIIC not only causes death or prolongs hospital stays but also worsens long-term survival [6]. This finding has been verified by many studies worldwide [7, 8]. Therefore, surgeons need to particularly pay attention to prevent the PIIC after gastrectomy.

Adequate preoperative evaluation on the risk assessment of complications is essential to decrease morbidity. Thus, it is known that many patient characteristics, such as age, comorbidity, and nutritional status are associated with mortality or morbidity [9,10,11]. A single factor is insufficient on risk evaluation; therefore, integral assessment tools including patient factors and intraoperative information, such as mE-PASS or POSSUM, had been devised [12, 13]. However, these indices have limitations. The mE-PASS poorly evaluates the patient’s general condition because only three comorbidities and no already well-known risk factors, such as sex and obesity, were included. POSSUM does not include comorbidities and cannot be assessed preoperatively because of the inclusion of intraoperative findings. Furthermore, the endpoints of these indices are mortality and overall morbidity. Therefore, their validity to assess risk for PIIC is unclear.

NCD is a major platform for the nationwide web-based surgical data entry system in Japan. The Japanese Society of Gastroenterological Surgery (JSGS) registry started to collect data in 2011 and has accumulated over 1.2 million gastroenterological surgery records from over 5000 hospitals throughout Japan. The endpoints of previous research in gastric cancer using the NCD were mainly mortality or overall morbidity, and PIIC was not yet recognized [2, 3, 14,15,16,17,18,19,20,21].

Therefore, this study aimed to establish a clinically available preoperative risk model for postgastrectomy PIIC for patients with gastric cancer. Here we employed the NCD system.

Methods

Data source

We used data from the JSGS registry. The NCD runs and maintains the database. All data variables, which were collected via a web-based platform, and their definitions are accessible on the NCD website (https://www.ncd.or.jp/, in Japanese). Considering that the registered data are used for board certification by JSGS and also by Japan Surgical Society, almost all major surgical procedures performed in Japan are expected to be recorded in this national database [22, 23]. The database committee conducts data verification through sight visit audits at randomly selected institutions [24].

The Ethics Committee of The Japanese Gastric Cancer Association (JGCA) approved this study. All patients were informed about the JSGS registry by the hospitals.

Study population

Figure 1 displays the study flow chart. Between January 2013 and December 2016, the NCD registered 198,934 patients who underwent distal or total gastrectomy for gastric cancer. Among these patients, 277 with missing data on age, 24 with missing body height data, 13,409 with metastatic gastric cancer, and 1288 who underwent emergent surgery were excluded. A total of 183,936 patients were finally included. Then, we divided them into development cohort (registered between 2013 and 2015, n = 140,558) and validation cohort (registered in 2016, n = 43,378).

Fig. 1
figure 1

Flow chart of the study population

Full size image

Endpoint

The incidence of PIIC was the primary endpoint of this study. Any morbidity that developed within 30 postoperative days indicated a postoperative complication. PIIC includes anastomotic leakage, intra-abdominal abscess, and pancreatic fistula. Pancreatic fistula was evaluated by the ISGPF criteria, and grade B or C was defined as pancreatic fistula [25].

Patient characteristics and predictors

Based on the potential patient data, tumor and surgical characteristics collected in the NCD-JSGS registry, the following 18 preoperative factors were selected by the research-promoting committee in JGCA as clinically important potential predictors of the outcome: age; sex; body mass index (BMI); activities of daily living (ADL); 12 comorbidity type such as diabetes mellitus, chronic obstructive pulmonary disease (COPD), hypertension, congestive heart failure, coronary artery disease, peripheral arterial disease (PAD), cerebral vascular disease, steroid use, anticoagulant therapy, chronic kidney disease, hemodialysis, and low albumin; surgical TNM cancer stage; and surgical procedure (total or distal gastrectomy). Myocardial infarction, angina, and coronary stent placement defined coronary artery diseases. Meanwhile, a less than 50 mL/min in creatinine clearance indicated chronic kidney disease. Most factors were preoperatively obtained, but the preoperative clinical cancer stage was substituted with surgical stage according to the intraoperative findings because no data were available on the clinical stage.

Statistical analysis

Using logistic regression, we constructed a multivariable risk model based on the development cohort with the 18 abovementioned explanatory variables to predict the occurrence of PIIC after gastrectomy in a primary analysis. The accuracy of the prediction was then tested in the validation cohort by using the c-statistics for discrimination and by calibration plot. Furthermore, we fitted another model that could be easily used in clinical practice, with a fewer number of variables in a secondary analysis. To create a parsimonious model with a reduced number of variables, we used the least absolute shrinkage and selection operator (LASSO) regression method. We chose the shrinkage parameter to eventually reduce the number of variables to 12. We scored the variables according to the odds ratios estimated in the model, excluded four variables, and finally narrowed to eight variables. The cohort was stratified into three risk groups for PIIC according to the score of the eight variables. Furthermore, we verified whether the groups were distinct in terms of the risks for PIIC in the validation cohort.

Results

Risk profile of the study population

Patients’ demographic data and cancer factors are summarized in Table 1. The population in the development cohort had a mean age of 70.5 years, and two-thirds of patients were males. Patients who were obese with a BMI of over 25 kg/m2 were around 20% of the development cohort. Major comorbidities were hypertension, followed by chronic kidney diseases, low serum albumin, and diabetes mellitus. Almost half of the patients showed stage I. Moreover, around 70% and 30% of patients underwent distal gastrectomy and total gastrectomy, respectively. Each prevalence of these descriptive data was similar between the development and validation cohorts.

Table 1 Key descriptive data in the development and validation cohorts
Full size table

Morbidity

All complications with grade II or higher in Clavien–Dindo classification occurred in 15.6% of the whole study population [26, 27]. The overall incidence of PIIC was 4.7% (Supplementary Table 1). In particular, 2.6% had anastomotic leakage, 1.7% had a pancreatic fistula, and 1.3% had intra-abdominal abscess.

Model results

The multivariate logistic model for PIIC in the development cohort is presented in Table 2. Age, sex, BMI, ADL, 12 types of comorbidities, gastric cancer stage, and surgical procedure were associated with morbidity incidence. Among them, male, obesity, peripheral vascular disease, steroid use, and total gastrectomy were strong factors. The predicted morbidity was based on the following logistic regression equation:

$${\text{Predicted morbidity }} = {\text{ e}}\left( { - {4}.{272 } + \, \Sigma \beta i} \right)/\left\{ {{1 } + {\text{ e}}\left( { - {4}.{272 } + \, \Sigma \beta i} \right)} \right\}$$
Table 2 Risk factors associated with postoperative intra-abdominal infectious complications in the development cohort
Full size table

where βi is the coefficient of the variable factors (Table 2).

Model performance

To evaluate the performance of this model, we performed calibration and discrimination by using the validation cohort. According to the calibration plot, the predicted probability and actual incidence fairly correlated (Fig. 2a). In addition, the area under the ROC curve (AUC) was 0.666 (95% confidence interval [CI], 0.660–0.672) (Fig. 2b).

Fig. 2
figure 2

Calibration and discrimination for the risk model. a Calibration plot between the development and validation cohorts. b Discrimination plot between the development and validation cohorts. The AUC was 0.666

Full size image

Simple prediction for intra-abdominal Infectious complication after Gastrectomy for gastric caNcer (SIGN score)

A reduced-risk model with simple and broad practicality for PIIC was constructed in the development cohort by using the LASSO regression method. Table 3 shows the estimates from the LASSO multivariable logistic model, which reduced the number of risk factors from 18 to 12. This reduced-risk model was expressed through the following equation:

$${\text{Predicted morbidity }} = {\text{ e}}\left( { - {3}.{799} + \Sigma \beta i} \right)/\left\{ {{1 } + {\text{ e}}\left( { - {3}.{799 } + \Sigma \beta i} \right)} \right\}$$
Table 3 Reduced-risk model for the postoperative intra-abdominal infectious complications in the development cohort
Full size table

These factors were scored to predict morbidity according to the coefficient (0.03 < β ≤ 0.09, 1 point; 0.09 < β ≤ 0.19, 2 points; 0.19 < β ≤ 0.29, 4 points; 0.29 < β ≤ 0.39, 6 points; 0.39 < β ≤ 0.49, 8 points; and 0.49 < β ≤ 0.59, 10 points). Four factors, namely, age, diabetes mellitus, COPD, and anticoagulant therapy, with a coefficient of less than 0.03 were deleted, and the scoring system eventually included eight factors and was named SIGN score. Male, obese, advanced cancer stage, and total gastrectomy strongly affected morbidity incidence. Patients were stratified into three risk groups (low-risk group, 0–7; middle-risk group, 8–15; and high-risk group, 16–31) according to the SIGN score. The predicted probabilities for PIIC in the low-, middle-, and high-risk groups were 1.8%, 4.1%, and 7.8%, respectively, using the development cohort. The prediction had good calibration against the observed incidence in the validation cohort (Fig. 3). The low-, middle-, and high-risk groups of the validation cohort had 9,917; 22,120; and 11,341 patients, respectively.

Fig. 3
figure 3

Comparison between the development and validation cohorts for the incidence of postoperative intra-abdominal infectious complication according to three risk groups. Left white bar indicates the predicted probability from the reduced-risk model in the development cohort. Right gray bar indicates the actual incidence in the validation cohort

Full size image

Discussion

In this study, a satisfactory risk model for PIIC was constructed using a large-scale database in clinical practice. As aging progresses worldwide, surgical high-risk patients also increase in number, and the importance of risk assessment further grows [28]. PIICs can be fatal, especially in fragile patients. Super-elderly individuals or those with multiple severe comorbidities were conventionally considered frail for surgery. Nonetheless, they can be evaluated through this risk model and may undergo limited gastrectomy or lymphadenectomy.

Risk stratification models for operative morbidities after gastrectomy have been constructed individually in distal and total gastrectomies by using the data from NCD [2, 3]. However, integral outcomes, such as PIIC, have not been investigated. To unify in a single model, we did not separate the distal and total gastrectomies. Although total gastrectomy was the strongest factor, the unified model remained robust.

Among the postgastrectomy complications, PIIC is a critical complication depending on surgical technique and is related to in-hospital death, long hospital stays or worsened long-term survival [29, 30]. The major reasons of death within 90 days after gastrectomy were cardiac events followed by PIIC, pneumonia, and events in the central nervous system in the NCD [15]. Intra-abdominal abscess mainly includes anastomotic leakage and pancreatic fistula; thus, we encompassed these three complications. In our study, the overall incidence of anastomotic leakage was 2.6%, which was the highest among the three complications. High leakage incidence may be caused by total gastrectomy, accounting for 30%. Anastomotic leakages in distal and total gastrectomies from the NCD in 2011–2012 were 2.1% and 4.5%, respectively [2, 3]. According to both retrospective and prospective cohort studies using NCD, pancreatic fistula frequently occurs in patients undergoing laparoscopic distal gastrectomy than open distal gastrectomy [17, 18]. Therefore, surgeons need to consider high-risk patients based on our model when performing laparoscopic distal gastrectomy.

Various types of comorbidity were associated with postoperative complications [31, 32]. This study revealed that preoperative steroid use, PAD, COPD, and hypertension strongly correlated with PIIC. Chronic steroid use increases infection risks or interferes with healing, and it is linked to anastomotic leakage [33]. Arteriosclerotic diseases such as hypertension or PAD causes negative effects on microvascular and oxygen perfusion and inhibit anastomotic healing [34, 35]. Low pulmonary function also leads to tissue hypoxia and anastomotic leakage [36]. Preoperative inflammation markers were not included in our model, because white blood cell and C-reactive protein with mild to moderate increase were not a major risk for abdominal abscess and pancreas fistula in previous NCD studies [2, 3].

Reduced-risk model can be available in general practice. However, despite its rigorous nature, considerably many factors or complicated calculations may disturb its routine use. Therefore, we narrowed 18 factors in the entire model to eight factors according to the prevalence and β coefficient to create a new scoring system, named as the SIGN score. It could stratify patients into three risk categories. Despite having high prevalence, factors with low β coefficient, such as diabetes mellitus, were excluded in this scoring system. Meanwhile, age was not an independent factor of PIIC in the reduced-risk model, and it supports the results of previous studies to indicate the feasibility and safety of gastrectomy in advanced age [37, 38]. Hence, limited surgery may be dispensable for healthy elderly patients.

This study has several limitations. NCD, which was used in this study, lacks detailed preoperative tumor information such as tumor location, size, clinical cancer stage, and planning of the surgical procedure and the extent of lymphadenectomy. Preoperative cancer stage is unknown, however, clinical stage includes a surgical stage in the 8th edition of the TNM classification [39]. Second, the severity of anastomotic leakage and intra-abdominal abscess is unknown. Third, neoadjuvant chemotherapy for gastric cancer is not a standard treatment in Japan. Only 3.4% of the study cohort received neoadjuvant chemotherapy. Therefore, we did not include it as a risk factor in this model, although neoadjuvant chemotherapy increases morbidity [40]. The strength of this risk model is that relevant risk factors for a wide range of comorbidities were identified and quantified using the nationally representative data. Some of the comorbidities could be controlled by preoperative intervention. For example, a preoperative physical exercise program could be implemented to control body weight and to build muscle, and preoperative respiratory rehabilitation could be proposed for patients with COPD [41]. Future challenges include evaluation of the efficacy of such preoperative intervention in reducing the incidence of PIIC in patients at high risk in this model.

Conclusion

A preoperative entire risk model for PIIC that occurs after gastrectomy for gastric cancer was developed on the basis of real-world nationwide clinical practice data. This led to the establishment of a new, simple scoring system that is more practical for clinical use. Other known risk factors had been included in the original risk analyses, and the identification of the most relevant eight factors is new to the literature. Thus, we believe that the current series of analyses, starting from the entire risk models including a wide range of comorbidities, was a valuable approach in establishing a novel tool for preoperative risk evaluation.