Risk model for morbidity and mortality following liver surgery based on a national Japanese database

Abstract Aim We evaluated the morbidity and mortality associated with liver surgery in Japan and developed a risk model for liver resection using information from a national database. Methods We retrospectively reviewed 73 861 Japanese patients who underwent hepatectomy between 2014 and 2019, using information from the National Clinical Database (NCD) registrations. The primary endpoints were 30 days and in‐hospital mortality, and the secondary endpoints were postoperative complications. Logistic regression risk models for postoperative morbidity and mortality after hepatectomy were constructed based on preoperative clinical parameters and types of liver resection, and validated using a bootstrapping method. Results The 30‐day and in‐hospital mortality rates were 0.9% and 1.7%, respectively. Trisectionectomy, hepatectomy for gallbladder cancer, hepatectomy for perihilar cholangiocarcinoma, and poor activities of daily living were statistically significant risk factors with high odds ratios for both postoperative morbidity and mortality. Internal validations indicated that the c‐indices for 30‐day and in‐hospital mortality were 0.824 and 0.839, respectively. Conclusions We developed a risk model for liver resection by using a national surgical database that can predict morbidity and mortality based on preoperative factors.

Program) database, which produces results of this nature across various surgical fields, including risk analysis of liver resection. 2,3The National Clinical Database (NCD) in Japan was developed in collaboration with the NSQIP, with the common goal of creating a standardized surgical database for quality improvement.[6][7][8][9] Analyses of risk models for liver resection that can be applied nationally have also been performed to date using NCD data. 10,11Notably, however, these analyses have been based on 1 or 2 years of data and are limited to certain types of liver resection in the study population.
Therefore, there is a need for longer-term and more comprehensive analyses.
In the present study, we examined more than 70 000 liver resections over a 6-year period at centers across Japan and performed a risk analysis of these operations based on information from the NCD.We further developed risk models for hepatectomy to improve the quality of these procedures.This study is the largest in the world to date on risk modeling of liver resection.

| Data collection
This data used in this study were obtained from the NCD, a nationwide collaborative project established in April 2010, involving the Japanese surgical board certification system.More than 5500 facilities participate in the NCD, and information on approximately 1.5 million cases is collected annually.The NCD is currently the world's largest clinical database linked to a medical specialty system.We retrospectively collected data from the NCD for liver resections conducted between 2014 and 2019.These cases included partial liver resection, segmentectomies 1-8, left lateral sectionectomy, right anterior sectionectomy, right posterior sectionectomy, right hepatectomy, left hepatectomy, central bisectionectomy, right trisectionectomy, left trisectionectomy, hepatectomy for gallbladder cancer, and hepatectomy for perihilar cholangiocarcinoma.All hepatectomies for gallbladder cancer or perihilar cholangiocarcinoma were classified as either hepatectomy for gallbladder cancer or hepatectomy for perihilar cholangiocarcinoma, even if a right or left hepatectomy, or a right or left trisectionectomy, was performed.After excluding duplicate procedures for the same patient, cases with discrepancies in the surgical method and site of liver resection, and cases with missing endpoints and covariates, 73 861 patients were finally included in the present study population (Figure 1).
The Ethics Committee of the NCD approved the retrospective use of data collected by the NCD for observational research.

| Endpoints
The primary endpoints of this study were 30-day mortality and inhospital mortality, defined as death within 30 days of surgery, regardless of whether the patient was in hospital or out of hospital, and death within the period of hospitalization, including 30-day mortality, respectively.Secondary endpoints included postoperative complications, such as reoperation, surgical site infection (organ space), bile leakage, pneumonia, renal failure, postoperative blood transfusion, and sepsis.

| Statistical analysis
Descriptive statistics were generated for the demographic, clinical, and laboratory characteristics of the study subjects.For continuous variables, the median and the 25th and 75th percentile values were calculated.Categorical variables were expressed as numbers and percentages.Risk models were developed for each outcome using backward stepwise logistic regression with Akaike's Information Criterion (AIC).Candidate covariates for inclusion in the risk models were selected from registry variables related to outcomes according to previous studies and clinical findings. 10,11Regression coefficients and intercepts, odds ratios (ORs), 95% confidence intervals (CIs), and P values for each coefficient were calculated for the final logistic regression model.The performance of the risk model was evaluated using a c-index and a calibration plot.Optimism of the c-index was corrected using Harrell's bias correction method.CIs for the c-index were calculated using the location-shifted bootstrap method because the sample size was sufficiently large. 12The number of bootstrap samples was set at 200.Bootstrapping was used to correct for optimism in the calibration plot.The statistical significance level was set at 0.05, and all tests were two-tailed.
All statistical analyses were performed using R software version 4.1 and later (R Foundation, Vienna, Austria, https:// www.

F I G U R E 1 Study population flow chart.
r-proje ct.org/ ).The backward stepwise variable selection method and calibration plots were performed using the rms R package version 6.2 and later (Frank E Harrell Jr., https:// CRAN.R-proje ct.org/ packa ge= rms).

| Risk profile of the study population
The clinical features of the study cohort are presented in Table 1.
The liver resections in our current series included partial and anatomical resections, such as segmentectomy, one sectionectomy, two sectionectomy, trisectionectomy, hepatectomy for gallbladder cancer, and hepatectomy for perihilar cholangiocarcinoma.The top three primary diagnoses included in this study were hepatocellular carcinoma in 30 093 (40.7%) patients, metastatic liver tumor in 22 726 (30.8%) patients, and intrahepatic cholangiocarcinoma in 5748 (7.8%) patients.The abbreviated risk profiles and preoperative laboratory data for the study population are shown in Table 1.Briefly, 12.4% of the patients had a history of preoperative chemotherapy within 90 days prior to surgery, 1.8% had a history of weight loss >10%, and 3.8% had a body mass index >30 kg/m 2 .The identified pre-existing comorbidities were as follows (frequency in parentheses): diabetes mellitus (internal medicine therapy) (17.1%), diabetes mellitus (insulin therapy) (5.9%), heavy alcohol use (30.5%), chronic obstructive pulmonary disease (3.9%), hypertension (40.6%), preoperative dialysis (0.9%), previous cerebrovascular disease (3.4%), chronic steroid use (1.0%), and anticoagulant therapy (6.0%).The American Society of Anesthesiologists (ASA) classification of III to V was assigned in 13.2% of the cases.

| Morbidities and mortalities after liver resection
The morbidity and mortality rates in the study cohort are shown in Table 2.The 30-day mortality rate was 0.9%, and the in-hospital mortality rate was 1.7%.With respect to postoperative complications, the rate of reoperation was 2.8%, surgical site infection was 5.0%, bile leakage was 7.3%, pneumonia was 1.7%, renal failure was 1.7%, postoperative blood transfusion was 4.3%, and sepsis was 2.0%.

| Risk models for 30-day mortality and in-hospital mortality
Logistic regression risk models for postoperative mortality after hepatectomy were constructed based on the preoperative clinical parameters and types of liver resection.Table 3 presents the results.
The mortality risk was calculated as follows: TA B L E 1 Clinical characteristics of the entire study cohort.

| Risk models for morbidities
Logistic regression risk models for postoperative morbidities were also constructed based on preoperative clinical parameters and types of liver resection and are presented in Table 4. Morbidity was calculated in the same way as mortality.With regard to postoperative complications, hepatectomy procedure was a statistically significant risk factor with a high odds ratio, as was mortality.Our present analyses further showed that, as well as risk factors for hepatectomy-related mortality, trisectionectomy, hepatectomy for gallbladder cancer, or hepatectomy for perihilar cholangiocarcinoma could be risk factors for a variety of postoperative complications with odds ratios of three or more times.Patient-side risk factors with odds ratios of three or more were ADL full support within 30 days of surgery, preoperative pneumonia for pneumonia, preoperative sepsis for renal failure, and preoperative sepsis for sepsis.

| Model performance
The model was validated internally using bootstrapping.The cindex and 95% CI for each outcome are shown in Table 5.The calibration curves of the models are shown in Figure 2, which also shows the values for 30-day mortality, in-hospital mortality, reoperation, surgical site infection (organ space), bile leakage, pneumonia, renal failure, postoperative blood transfusion, and sepsis.
These calibration curves indicated how well the predicted event rates match the observed rates among patient risk groups.The calibration plot showed that, for many models, the line deviates from the ideal straight line, except where the prediction probability is low, and overestimates where the prediction probability is high, as indicated in Figure 2.For outcomes with a particularly low number of events, such as 30-day mortality and in-hospital mortality, the calibration plot deviates from the ideal straight line, except where the calibration plot has a low predicted probability.
However, for outcomes with a relatively high probability of occurrence, such as biliary leakage, the calibration plot was in good agreement with the ideal straight line.In contrast, the c-indices for

TA B L E 3 (Continued)
30 days mortality and in-hospital mortality were 0.824 and 0.839, respectively, indicating good performance with respect to the outcomes with a low number of events.

| Thirty-day mortality and in-hospital mortality by number of cases per facility
Figure 3a shows the average number of cases per facility in each year.Figure 3b,c show the 30-day and in-hospital mortality rates by facility size, respectively.We examined the 30-day and in-hospital mortality rates by institution size (i.e. less than 30 cases per year, 30-49 cases per year, or 50 cases or more per year) based on the number of liver resections performed per year (Figure 3b,c).As shown in Figure 3b,c, facilities that conducted 30-49 and 50 or more hepatectomies per year tended to have lower 30-day and in-hospital mortality rates than those that had fewer than 30 hepatectomies in any year over the 2014 to 2019 study period.In a comparison of facilities with 30-49 against those with 50 or more liver resections per year, there were years in which the 30-day and in-hospital mortality rates were equivalent and years when this rate was lower in facilities that had conducted 50 or more procedures.

TA B L E 4
Risk models for postoperative morbidities.

| DISCUSS ION
Here, we examined more than 70 000 liver resections performed at Japanese centers over a 6-year period and analyzed the risks associated with a liver resection procedure, based on a national Japanese database.In our analyses, the 30-day mortality rate was 0.9%, and the in-hospital mortality rate was 1.7% in the study population.These values were lower than those described in previous NCD reports from Japan. 10,11Furthermore, the mortality rates obtained in our current study are lower than those reported by other national databases. 2,3This indicated that good surgical standards were maintained in Japan during the study period.We also developed a risk model for liver resection using the Japanese NCD with good-quality assurance.
We observed that the statistically significant risk factors with an odds ratio of three times or more for 30-day mortality in our study population were trisectionectomy, hepatectomy for gallbladder cancer, and hepatectomy for perihilar cholangiocarcinoma.Those for inhospital mortality were right hepatectomy, right trisectionectomy, left trisectionectomy, hepatectomy for gallbladder cancer, and hepatectomy for perihilar cholangiocarcinoma.
Trisectionectomy and hepatectomy for perihilar cholangiocarcinoma had the highest odds ratios.These variables have been reported as risk factors for liver resection in previous studies.Lang et al. reported an operative mortality rate of 11.9 % associated with left trisectionectomy. 13Farid reported a 90-day mortality rate from left trisectionectomy of 9.7%. 14Recently, Kron et al. reported an overall 90-day mortality rate of 7.6 % from trisectionectomy for hepatopancreatobiliary malignancies. 15A high perioperative mortality has been similarly reported following surgery for perihilar cholangiocarcinoma, for which Mueller et al. reported median inhospital and 3-month mortality rates of 4.7% and 7%, respectively. 16cently, the Perihilar Cholangiocarcinoma Collaboration Group reported a 90-day mortality rate of 13.6% for perihilar cholangiocarcinoma using an international database. 17Perioperative mortality rates >10% for perihilar cholangiocarcinoma have been reported elsewhere. 18Similarly, 90-day mortality rates of more than 10% have been reported for gallbladder cancer. 19e risk factors for perioperative mortality in liver surgery identified in the present study are consistent with previous reports.In contrast, full ADL support within 30 days prior to surgery was identified as a risk factor for both 30-day mortality and in-hospital mortality.
In addition to the liver resection protocol, poor capacity to tolerate surgery can also increase the risk of surgery-related death.1][22] Our current results are consistent with those of previous reports on patient-related risk factors.Collectively, these findings indicate that massive major hepatectomy, liver resection with biliary reconstruction, and assisted ADL should be recognized as potential risk factors for hepatectomy-related death.The present study was also based on information from a national database and an extremely large sample size, which contributes to the reliability of the findings.
Here, we analyzed the risk factors not only for operation-related mortality but also for postoperative complications that arise following liver resection.Regarding postoperative adverse events, the type of liver resection was a statistically significant risk factor, with a high odds ratio.Similar to mortality, trisectionectomy, hepatectomy for gallbladder cancer, and hepatectomy for perihilar cholangiocarcinoma were identified as risk factors of various postoperative complications in the present study.Farid reported that postoperative morbidity arising from left trisectionectomy ranges from 45% to 50%. 14Kron et al.
found that 40.3% of patients who underwent right trisectionectomy had postoperative complications. 15Mueller et al. reported an overall postoperative complication rate of 80.5% after surgery for perihilar cholangiocarcinoma, with severe complications (grade ≥IIIa) occurring in 58.1% of these patients. 16Other hepatectomy procedures with a large liver dissection area, such as right anterior sectionectomy and central bisectionectomy, were also observed in the present study cohort to be a risk factor for bile leakage and other adverse events.
Other prior studies have also demonstrated that right anterior sectionectomy and central bisectionectomy are risk factors for bile leakage. 23,24On the other hand, the patient-side risk factors identified herein for postoperative complications with odds ratios of three or more times were full ADL support within 30 days before surgery, preoperative pneumonia, and preoperative sepsis.

TA B L E 4 (Continued)
The C-indices of 30-day mortality and in-hospital mortality in this present study were 0.824 and 0.839, respectively (Table 5).The C-indices of bile leakage, pneumonia, renal failure, and sepsis in this study were 0.739, 0.750, 0.803, and 0.774, respectively ( Our current results thus appear to have superior prediction accuracy for both morbidity and mortality except for sepsis.It must be noted of course that the validity of this comparison is impacted by differences in the data analyzed.However, our present results suggest that the risk model we have formulated in this current study for both morbidity and mortality may be reliable in clinical practice. We found from our current analyses that the hospital volume affected both the 30-day and in-hospital mortality.It is known that the degree of correlation between hospital volume and surgical mortality varies significantly among surgical procedures, and the concept of "failure to rescue" has recently gained attention with regard to postoperative mortality. 25,26Postoperative morbidity rates are similar between hospitals with low mortality and those with high mortality, but there is a significant difference in the success rate of postoperative management (i.e., "failure to rescue"), which is believed to be related to mortality rates.It should be recognized therefore that in addition to the preoperative factors examined in this study, "failure to rescue" can also influence mortality.
The present study had some notable limitations.First, it was limited to the Japanese population.Validation using databases from other countries is necessary to evaluate the general applicability of the results.Second, although we analyzed the procedures used for liver resection in our large cohort, we did not stratify these cases by open or laparoscopic surgery.To ensure better comparisons between the centers represented by our study population and to try and more fairly reflect the outcomes, our approach was to assess risk in accordance with patient background factors and not to include surgeon-selectable factors or perioperative events that were not known preoperatively among the covariates.In addition, we did not include institutional factors in the risk model variables.Previous reports have shown that center effects can also influence outcomes in the field of hepatobiliary surgery. 27,28We here evaluated the risk of all hepatectomies based on patient background factors alone.The reason for this was that our risk model aimed to compare outcomes at different centers based solely on preoperative patient factors.
The inclusion of center effects as a predictor variable seems to be contrary to a fair comparison between facilities.It should be recognized, however, that center effects may influence outcomes.Third, for outcomes with a small number of events, the predicted probability tended to overestimate the actual probabilities.Hence, surgeons should be aware of the possibility of overestimation of outcomes with a low number of events.
In conclusion, we examined more than 70 000 Japanese liver resection cases using a nationwide surgical database and developed a risk model for these operations.This model can predict 30-day mortality, in-hospital mortality, and major and life-threatening postoperative morbidities after liver resection, based on preoperative factors alone.Hence, it can assist surgeons and patients in better understanding the risks of these surgeries and making appropriate preoperative decisions.

FU N D I N G I N FO R M ATI O N
This work was supported by the Japanese Society of Gastroenterological Surgery.

CO N FLI C T O F I NTE R E S T S TATE M E NT
Hiroaki Miyata is affiliated with the Department of Healthcare Quality Assessment at the University of Tokyo.The department is a social collaboration department supported by grants from the National F I G U R E 3 (A) Average number of cases per facility in each year.(B) 30-day mortality rates by facility size.(C) In-hospital mortality rates by facility size.Clinical Database, Johnson & Johnson K.K, Nipro Corporation, and Intuitive Surgical Sàrl.Shinya Hirakawa and Hisateru Tachimori belong to an endowed course of Keio University funded by Takeda Pharmaceutical Company Limited and belong to the Department of Healthcare Quality Assessment at the University of Tokyo that accepts financial support from National Clinical Database, Johnson & Johnson K.K., Nipro Corporation, and Intuitive Surgical Sàrl.But Dr Miyata, Dr Hirakawa, and Dr Tachimori have no conflicts of interest regarding this research.The other authors declare no conflicts of interest regarding this research.Dr. Yoshihiro Kakeji and Dr. Ken Shirabe are editorial members of Annals of Gastroenterological Surgery.E TH I C S S TATEM ENTS The Ethics Committee of the NCD approved the retrospective use of data collected by the NCD for observational research.Informed Consent: N/A Registry and the Registration No. of the study/Trial: N/A Animal Studies: N/A O RCI D Tatsuya Orimo https://orcid.org/0000-0003-4398-0697Taro Oshikiri https://orcid.org/0000-0003-0635-3432Yoshihiro Kakeji https://orcid.org/0000-0002-2727-0241R E FE R E N C E S

chronic obstructive pulmonary disease. TA B L E 1 (Continued) TA B L E 2
Morbidities and mortalities among the NCD hepatectomy study population (n = 73 861).

Overall incidence, n (%)
TA B L E 3 Risk models for 30 days mortality and in-hospital mortality.TA B L E 3 (Continued) (Continues)Abbreviations: ADL, Activities of daily living; ASA, American Society of Anesthesiologists; CI, confidence interval; COPD, chronic obstructive pulmonary disease.