Risk assessment of morbidities after right hemicolectomy based on the National Clinical Database in Japan

Abstract Objective Nationwide databases are expected to provide critical data to improve medical practice. The present study used such data to develop risk models for clinically important outcomes after right hemicolectomy based on preoperative risk factors. Methods Japan's National Clinical Database (NCD) identified 38 030 cases of right hemicolectomy in the years 2011 and 2012. These were used to analyze correlations between mortality and eight selected clinical outcomes of interest by Pearson's correlation coefficient (r). To construct risk models for the eight selected clinical outcomes, 80% of all the examined cases were extracted randomly as a development cohort, and preoperative risk factors for each clinical outcome were identified using a forward stepwise selection method. Morbidities predicted from the risk models were used to find areas under the receiver operator curves among the remaining 20% of the testing cohort. Results The following clinical outcomes were identified as highly associated with operative mortality: systemic sepsis (r = .360), renal failure (r = .341), unplanned intubation (r = .316) and central nervous system (CNS) occurrences (r = .301). Risk models containing up to 21 preoperative variables were constructed for these eight postoperative clinical outcomes. Predictive values of the eight models were as follows: surgical site infections (0.634), anastomotic leakage (0.656), systemic sepsis (0.816), pneumonia (0.846), unplanned intubation (0.838), renal failure (0.883), CNS occurrences (0.833) and transfusion >5 units (0.846). Conclusions This study indicated that the NCD‐generated risk models for six of the eight selected critical postoperative outcomes had high discrimination among right hemicolectomy patients. These risk models can accurately identify high‐risk patients prior to surgery.

This program demonstrably enhanced the quality of surgical care. 2 In Japan, the National Clinical Database (NCD) was founded in April 2010 as the parent body of the database system linked to the board certification system. Registration began in 2011 and, to date, more than 4600 facilities have enrolled and more than 1.5 million cases are being recorded each year. Nationwide data collection and analysis leads to high-quality health care for patients and the general public. 3 Since 2014, the NCD has provided a preoperative risk calculator tool that allows clinicians to predict an individual patient's mortality risk for eight surgical procedures. [4][5][6] Analysis of 19 070 right hemicolectomy cases registered in 2011 indicated that 30-day mortality (death within 30 days after surgery in or out of hospital) was 1.1% and operative mortality (death within 30 days of surgery or within 90 days of surgery during the same hospitalization) was 2.3%. We previously used the NCD dataset to examine risk models for 30-day mortalities and operative mortalities after right hemicolectomies; 7 however, risk models for postoperative morbidity except for operative mortality have not been evaluated until now.
Herein, we report the development and performances of risk models for eight clinical outcomes associated with high mortality or having high incidence after right hemicolectomy, predicting morbidity rate based on preoperative risk factors, using the NCD data from 2011 and 2012.

| Data source
Japan's National Clinical Database continuously manages registry data by a web-based data management system, and identifies individuals with annual data approval. We carried out this study using the Japanese Society of Gastroenterological Surgery (JSGS) registry data in the NCD. We enrolled patients who underwent right hemicolectomy between 1 January 2011 and 31 December 2012. A total of 38 924 cases of right hemicolectomy were registered in the NCD during this time. As in our previous study, we excluded 894 cases that had missing values for basic data such as patient gender, age or 30-day mortality or those that had a simultaneous surgical procedure such as esophagectomy and hepatectomy. 7 After these exclusions, 38 030 cases that had all the parameters treated in the analyses remained in the study.
The types of data recorded into the JSGS registry in the NCD are almost identical to those used by the ACS NSQIP. Potential independent variables included 15 patient demographic variables, 46 pre-existing comorbidities, and 19 preoperative laboratory values.
Preoperative variables in the NCD are listed in

| Pneumonia
(occurrence within 30 days after surgery): It is defined as inflammation of the lung caused by bacteria, viruses or chemical irritants (it did not exist before surgery). It develops with cold sweat, fever, chest pain, cough and purulent sputum within 30 days after surgery.
Definition of pneumonia requires that one of the following two criteria must be met:  Definition of SIRS requires that two of the following criteria must be met: (i) body temperature >38°C or <36°C; (ii) heart rate >90/ minute; (iii) respiration rate >20/minute or PaCO 2 <32 mm Hg; (iv) leukocytes >12 000 cells/mm 3 or <4000 cells/mm 3 or bands >10%.

| Sepsis
Sepsis shows signs and symptoms of SIRS. Definition of sepsis requires that one of the following criteria must be met: (i) blood culture positive; (ii) proof of maturation or detection of bacterial culture from infectious lesion.

| Statistical analysis
Pearson's correlation coefficients between each of the eight complications and operative mortality were estimated. Following the definition used in previous studies from NCD, we defined 30-day mortality as death within 30 days regardless of hospitalization status, and operative mortality as death within 30 days or at discharge.
For risk modeling, we first divided the cohort randomly into an 80% development cohort and a 20% testing cohort, because building a model in a dataset with a large number of outcomes was more important than testing the accuracy (i.e. estimating c-statistics or depicting receiver operating characteristic [ROC] curves) of the model with greater precision. Within the development cohort, we constructed logistic regression models for the outcomes of interest from the preoperative risk factors listed above, using a forward stepwise selection method with an inclusion criteria of P < .05 and an exclusion criteria of P ≥ .10. We then applied the models to the patients in the testing cohort, predicting their baseline risk of the outcome using the following equation: Predicted morbidity was calculated as e ðb0þ P bixiÞ =1 þ e ðb P bixiÞ where bi is the coefficient of variable Xi from the logistic regression model in the development cohort. 4 Using the predicted risk, we assessed the discrimination of the models using ROC curves and their c-indices. We used IBM SPSS Statistics for Windows (Version 20; IBM, Armonk, NY, USA) for data analysis.

| RESULTS
The complication of "any SSI" was identified most frequently and its incidence rate was 7.5%. Morbidity rates of other clinically important postoperative outcomes were as follows: anastomotic leakage   Table 3.

| Model performance
Concordance indices (C-indices) of the risk models for these major postoperative morbidities estimated in the testing samples (n = 7423) are summarized in Table 4, and the resultant eight ROC curves are shown in Figure 1. The models for sepsis, pneumonia, Although the accuracy of prediction for SSI or anastomotic leakage did not reach the level of those for the morbidities associated with mortality, the models suggested the relation of certain preoperative comorbidities and laboratory data to the occurrences of SSI and anastomotic leakage, which hamper early discharge of patients and are a significant source of aggravation to surgeons and patients alike. We suspected that these results indicated that not only preoperative factors but also intraoperative factors influenced this risk model of SSI or anastomotic leakage such as non-life-threatening postoperative complications.

T A B L E 2 Correlation between mortality and the 8 selected postoperative morbidities
There have been various studies of risk stratification of mortality and/or morbidities using a nationwide database after surgery. 8,9 Recently, the Surgical Risk Preoperative Assessment System (SUR-PAS) showed that accurate preoperative risk assessment of postoperative mortality, overall morbidity, and six complication clusters in a broad surgical population could be achieved with as few as eight preoperative predictor variables. 9 However, the authors pointed out some limitations because as they cover a broad spectrum of operations carried out by nine surgical specialties (general, vascular, orthopedic, otorhinolaryngological, urological, thoracic, plastic, gynecological, and neurosurgery), generic predictors and outcome variables were necessarily chosen rather than more operation-or disease-specific ones. They also mentioned that these specific com-  [14][15][16][17] These effects could also be evaluated in Japanese patients by risk-adjusted analysis using the risk models created in this study. Based on the concept of the risk model, we can share the incident rates about postoperative complications with medical staff under intensive perioperative management, and we can also consider surgical procedure instead of right hemicolectomy in order to reduce the risks of mortality and morbidity. Furthermore, it is possible (but not for emergency patients) to prepare preoperative conditions on preoperative risk factors shown in Table 3.
Our study has some important limitations to consider. First, emergency surgery, especially for those in the septic condition, has great impact on the mortality and morbidity of patients. Mortality and morbidity are typically high in such patients compared to those undergoing elective surgery. In this analysis, patients who were diagnosed with acute diffuse peritonitis were categorized as critically ill patients undergoing surgery, and risk models for these patients were created separately as reported previously. 18 Second, the risk models created in the present study were based strictly on Japanese patients. In the previous collaborative study between NSQIP and NCD, we found that patient background, comorbidities, and practice style in Japan and the USA had key differences. 8 In the models, the OR for each variable was similar between NCD and ACS-NSQIP, but some risk predictors were population specific. The generalizability of these models to patients in other countries needs to be further investigated in future studies. Future cross-population evaluations are expected to help prevent complications and reduce health-care spending. 19 Risk models of postoperative outcomes in the present study were analyzed using only preoperative variables. In contrast, intraoperative factors influence postoperative outcomes. Risk assessment including intraoperative variables will be evaluated in a future study from the NCD.
In conclusion, we analyzed the performance of risk models for the eight most serious and/or common postoperative adverse events after right hemicolectomy. This study showed that the NCD-generated risk models, especially for the six lethal postoperative outcomes examined, worked well in discriminating patients who develop these events after right hemicolectomy surgeries.
These risk models will greatly enhance the evaluation of individual patients prior to surgery.