Development and validation of grade‐based prediction models for postoperative morbidity in gastric cancer resection using a Japanese web‐based nationwide registry

Abstract Aim Gastric cancer is the second leading cause of cancer death worldwide. Surgery is the mainstay treatment for gastric cancer. There are no prediction models that examine the severity of postoperative morbidity. Herein, we constructed prediction models that analyze the risk for postoperative morbidity based on severity. Methods Perioperative data were retrieved from the National Clinical Database in patients who underwent elective gastric cancer resection between 2011 and 2012 in Japan. Severity of postoperative complications was determined by Clavien‐Dindo classification. Patients were randomly divided into two groups, the development set and the validation set. Logistic regression analysis was used to build prediction models. Calibration powers of the models were assessed by a calibration plot in which linearity between the observed and predicted event rates in 10 risk bands was assessed by the Pearson R 2 statistic. Results We obtained 154 278 patients for the analysis. Prediction models were constructed for grade ≥2, grade ≥3, grade ≥4, and grade 5 in the development set (n = 77 423). Calibration plots of these models showed significant linearity in the validation set (n = 76 855): R 2 = 0.995 for grade ≥2, R 2 = 0.997 for grade ≥3, R 2 = 0.998 for grade ≥4, and R 2 = 0.997 for grade 5 (all: P < 0.001). Conclusion Prediction models for postoperative morbidity based on grade will provide a comprehensive risk of surgery. These models may be useful for informed consent and surgical decision‐making.


| INTRODUC TI ON
Gastric cancer is the fourth leading cause of cancer incidence and second leading cause of cancer death worldwide. 1 Surgery is the only treatment that provides a chance for a cure against gastric cancer except for endoscopically resectable early tumors. 2 Surgery for gastric cancer can be safely carried out in most cases. Thirtyday postoperative mortality rates of gastric cancer resections were <1% according to a national database in Japan. 3,4 Nevertheless, elderly patients who require gastric cancer resections are increasing with an aging society in developed countries. Some of these patients have multiple comorbidity conditions, such as hypertension, diabetes mellitus, old myocardial infarction, and old brain infarction, and patients sometimes go back and forth between nursing homes and hospitals. These patients have limited reserve capacity and sometimes suffer from postoperative complications.
Therefore, prediction of postoperative morbidity is still important.
There are several prediction models of postoperative morbidity for gastric cancer patients. [5][6][7][8][9][10] Nevertheless, patients may be uncertain about how much damage they will sustain because there are no prediction models based on severity. For example, patients can eat meals and function normally because of a wound infection. By contrast, patients are kept in bed with various tubes because of an abdominal abscess. To share information on postoperative morbidity, prediction of grade-specific morbidity rates is needed for both patients and doctors.
The National Clinical Database (NCD) in Japan was developed in collaboration with the National Surgical Quality Improvement Program (NSQIP) in USA with a shared goal of creating a standardized surgery database for quality improvement. NCD and NSQIP have developed systems using standardized variable definitions to collect data on risk factors and outcomes after surgery. 11 These databases collect prospective rather than retrospective data. Because patient registration for the board certification system by the Japan Surgical Society can be carried out only by the NCD system, the NCD now covers more than 97% of total surgical procedures in Japan. 12 This study was undertaken to construct prediction models to estimate grade-specific postoperative morbidity in gastric cancer resection using large NCD data.

| Study design
This study was prepared in response to a public call for research using NCD data by the Japanese Gastric Cancer Association (JGCA) in 2014.
The study protocol was approved by the Council of JGCA on June 3, 2014. This was a retrospective analysis of data from NCD data.

| Patients
Patients were selected who underwent partial gastrectomy, total gastrectomy, total gastrectomy with splenectomy, and total gastrectomy with distal pancreatectomy and splenectomy in combination with the main disease of gastric cancer between 2011 and 2012.
Exclusion criteria were emergency operations, concomitant cancer, preoperative systemic inflammatory response syndrome (SIRS), preoperative sepsis, and recurrent disease because standard operations are usually avoided under these conditions.

| Data collection
Thirty-eight preoperative conditions including comorbidities, past history, and functional status were collected with 17 preoperative laboratory data. Sixteen intraoperative data were also collected including type of surgery, American Society of Anesthesiologists Physical Status (ASA-PS), operative duration, blood loss, intraoperative problems, and information of metastatic organs. 11 The outcome of this study was postoperative morbidity according to the Clavien-Dindo classification. 13 Postoperative complications were defined as adverse events that occurred within 30 days after operation. Absence or presence of 47 complications was recorded with Clavien-Dindo grades. Other complications were also recorded with their name and Clavien-Dindo grade.

| Statistical analysis
Patients were divided into two groups, a development set and a validation set, by a random sampling method. Univariate analysis of each predictor was conducted using chi-squared tests with Yates correction when appropriate. Using the significant variables by univariate analysis, a stepwise increase logistic regression analysis was carried out to construct a prediction model for graded postoperative morbidity. When entering the variables into the multivariate analyses, we excluded the variables with ambiguous definitions, such as "transport by ambulance." We also excluded the variables with an incidence less than 1%, such as "ventilator dependent." We further excluded the variables with odds ratios that were close to 1 even though they were significant. When the variables were closely related to each other, such as hemoglobin levels and hematocrit, we excluded the variable with the smaller odds ratio.
Discriminative and calibration power of the model was carried out using area under receiver-operating characteristics curve (AUC) and a calibration plot, respectively. In the calibration plot, patients were divided into 10 risk bands according to the predicted event rates. Each risk band was set to have an equal number of patients.
Linearity between the observed and predicted event rates was assessed by Pearson's R 2 statistic.

| RE SULTS
We obtained 154 278 patient datasets for analysis and divided them into two groups, a development set and a validation set.
Demographic data are shown in Table 1 where β 0 is a constant, β i is a coefficient, and X i is a variable). "-" means no coefficients.
TA B L E 2 Coefficients of logistic regression models for postoperative morbidity Table 2 shows the parameters and coefficients of the grade-specific models for postoperative morbidity that were generated in the development set. The predicted event rate (p) for each outcome was calculated as follows: ln[p/(1 − p)] = β 0 + ∑β i X i , where β 0 is a constant, β i is a coefficient, and X i is a variable.   Tables 3 and 4 show the observed event rates in each risk band determined by the prediction models for grade ≥2 and grade ≥3, respectively.

| D ISCUSS I ON
We constructed and validated prediction models for postoperative morbidity following gastric cancer resections according to severity bias. 20 The current models were constructed by analyzing the variables of physiological variables, metastatic status, and type of surgery. In preliminary analyses, we also constructed prediction models by adding surgical factors. Nevertheless, the predictive power of models with surgical factors was similar to models without surgical factors. Therefore, we used models without surgical factors. should not be used for patients that need an operation. However, when we classified a group of patients using the system as shown in Tables 3 and 4, we can assume the risk of these patients. In contrast, the models for grade ≥4 and grade 5 showed good discriminative and calibration power. Therefore, we can directly use the predicted event rates for these patients. Using these models, physicians and patients can be informed of operation risks more specifically.
When using these models, the quality of care of hospitals should be considered because a volume outcome relationship has been reported for gastric cancer surgeries. [21][22][23][24] The present models provide an average risk in Japan. Before using the models, each hospital should calculate postoperative morbidity rates according to the Clavien-Dindo grading and compare them with NCD data. If the rates are far from NCD data, the predicted event rates should be adjusted in parallel to the observed rates.
A limitation of the present study was that the present validation set was not truly an external subset. We must analyze the predictive power in a truly external set such as future data. A validation study outside Japan will also add generalizability. Furthermore, the present models include many independent variables, which may burden the working time for data input. However, these variables are already incorporated into the NSQIP/NCD system and will not increase the workload at participating hospitals. The authors will upload a computer file which can compute risk of postoperative morbidity using the current models.
In conclusion, we constructed prediction models for grade-based postoperative morbidity using large national data. These models