In-hospital mortality after stomach cancer surgery in Spain and relationship with hospital volume of interventions

Stomach cancer is the second most common malignancy of the digestive tract in developed countries [1]. In Spain, the incidence adjusted to the worldwide population ranges from 12.2 to 21.6 cases per 100 000 men, depending on the region; the incidence in women is slightly less than half that of men. Surgery and chemotherapy are the mainstays of treatment. However, surgery is associated with considerable morbidity and lesser though significant mortality. The few studies published on morbidity and mortality after surgery for gastric cancer report variable rates [2‐4].

In Spain, gastric cancer surgery is performed in many types of hospitals and in all regions. On the other hand, there is no specific register that facilitates the assessment of process and outcomes of surgical interventions.

In-hospital mortality has often been considered an outcome indicator directly related with the quality of care [5]. Because in-hospital mortality is an objective measurement that is readily available in hospital databases, it has been used to analyze and compare outcomes among different centres. However, to ensure valid comparison, it is necessary to adjust the rates by taking patients' baseline risk or comorbidities into account [6, 7]; thus, different methods have been validated to be used with administrative databases with codes for diagnoses and procedures [8, 9]. In the absence of specific registers, administrative databases are the main alternative for this kind of evaluation.

In addition to patients' baseline condition, aspects related to the structure of the hospital, the experience of the professionals involved, and the surgical procedure itself can affect surgical outcomes. Likewise, a centre's volume of activity for a given type of surgical procedure, especially for cardiovascular and oncological interventions, has also been reported to affect post-operative mortality in several studies [10‐14]. However, some recent studies question the relationship between volume of activity and outcome; the authors of these studies point out that even if increased volume of activity were responsible for better outcome, the mechanisms underlying improved outcomes are not clear [13, 15, 16]. On the other hand, different definitions and cut-off points referring to hospital volume could be responsible for the divergent results found among different studies.

Given the relatively high rate of in-hospital mortality for gastric cancer reported by various authors, the scarcity of studies that analyze the surgical outcomes of this malignancy in Spain, and the controversies related to the possible association between volume of activity and outcomes, this study aimed to: 1. estimate the in-hospital mortality in surgery for gastric cancer in different regions in Spain; 2. identify factors associated to greater in-hospital mortality; 3. analyze the possible relation between volume and in-hospital mortality.

We performed a retrospective cohort study (based on administrative database) of all patients discharged after surgery for stomach cancer during 2001 and 2002 in four regions of Spain. These regions represent about 52% of the total population. In Spain, there is neither a common oncological surgical registry nor a National Cancer Registry. For many years, all hospital discharges are homogenously recorded and centralized at the Department of Health in each of the 17 Autonomous Communities or regions in the administrative database called Minimum Basic Data Set for Hospital Discharges (MBDS-HD). This database contains the following information: date of birth, gender (male or female), type of admission (urgent or scheduled), destination on discharge (dead or alive), International Classification of Diseases 9^th revision Clinical Modification (ICD9CM) [17] codes for the main and secondary diagnoses, ICD codes for the main and secondary procedures performed, date of admission, and date of discharge.

We included all discharges corresponding to patients with a principal diagnosis of stomach cancer (ICD code: 151.XX) that had undergone total or partial gastrectomy (ICD code: 43.5–43.9).

Secondary diagnoses were grouped into 259 mutually exclusive categories using the Clinical Classifications Software (CCS) [18] developed by the Center for Organization and Delivery Studies in the Healthcare Cost and Utilization Project (HCUP) at the Agency for Healthcare Research and Quality (AHRQ).

To pre-select factors that might be associated to in-hospital mortality, we contacted oncologists, gastroenterologists, and surgeons from different centres. We asked them to propose a list of surgical factors, patient comorbidities, factors related to the severity of disease, and complications that they considered might increase the probability of in-hospital death during or after surgery. The possible factors suggested and corresponding ICD codes are listed in Appendix 1. Although the stage of the tumour was among the factors proposed, it was not included in the study because the MBDS-HD does not include a specific code for this factor and no population cancer registry was available.

The study was approved by the institutional review board of the Corporació Sanitària del Parc Taulí.

Apart from the factors listed in the appendix, the following variables were considered: age group (≤50, 51–64, 65–74, 75–84, ≥ 85), gender, region, type of admission as recorded in the MBDS-HD (urgent or elective), and volume of discharges analyzed for each hospital. For each admission, the Charlson score was calculated from the codes for the secondary diagnoses using the Deyo [8] adaptation; each case was then grouped into one of four categories (0, 1, 2, > 2). We calculated the length of stay for each admission. We also created the variable 'number of secondary diagnosis coded' for each discharge, which was later recoded into the categories ≤ 3, 4–5, and ≥ 6.

In-hospital mortality was defined as death occurring during the hospital stay. The annual volume of discharges was defined as the mean number of discharges included in the study at a given centre per year. Annual volume of discharges was grouped into three categories according to terciles (<18, 18–35, >35) and into 7 volume categories corresponding to smaller ranges consisting of 10 discharges each.

The unit of analysis was the hospital discharge. We carried out a descriptive analysis of all variables of interest. The overall and specific in-hospital mortality rates for stomach cancer were estimated as a function of the admission type, age group, gender, region, annual volume of discharges, CCS diagnoses selected, and type of surgical procedure. The 95% confidence intervals were calculated for the overall rate according to the normal approximation. The chi-square or the Fisher's exact test was used to determine whether the factors studied were associated to mortality. Then, the same type of analysis was used to compare some variables of interest (age, gender, mortality, Charlson score, type of admission, region), as a function of the 3 annual volume categories. We used the Kruskal-Wallis test to compare the mean number of secondary diagnoses registered per discharge and the mean length of stay.

Then, a logistic regression model was constructed to determine whether the different demographic (age, region), admission factors (urgent, number of secondary diagnoses), or comorbidities studied (Charlson, congestive heart failure, pancreatic disorders, cardiac dysrhythmias, nutritional deficiencies, gastrointestinal haemorrhage, other gastrointestinal disorders, invasion of other structures) were independently associated to the adjusted mortality. Only those secondary diagnoses considered comorbidities by the experts and not included in the Charlson score were considered for the model, so possible complications occurring as a consequence of the intervention were not included (see appendix 1). First, we selected variables present in more than 1% of cases (more than 30 cases) that had p values < 0.1 in the univariate analysis. Next, we used the forward conditional stepwise method to construct the model. The odds ratios and 95% confidence intervals were calculated. Finally, goodness of fit was evaluated by the Hosmer-Lemeshow X² statistic [19] and the area under the receiver operator characteristic (ROC) curve was calculated to assess the discriminative capacity of the model. Values ranging from 0.7 to 0.8 represent reasonable discrimination and values exceeding 0.8 represent good discrimination [20].

We evaluated the association between hospital volume and adjusted mortality by introducing the variable annual hospital volume (3 categories) in the logistic regression model and estimating its odds ratios and 95% confidence intervals.

Differences in study periods and the definition of mortality used (such as post-operative mortality, 30-day mortality, or in-hospital mortality) among the different studies published limits the comparability of results. Moreover, some studies, such as ours, did not adjust mortality rates for severity factors, such as tumour stage at diagnosis. Despite these limitations, we can say that the in-hospital mortality rate observed in our study was high, although it was within the range of 1.7% to 12% reported by other authors [2, 21, 22]. McCulloch et al. reported the exact same mortality rate in 4 years as found in our study [23]. Furthermore, the wide range of variability among hospitals in our study might be partly due to differences in the factors that we found were associated, as the estimations of the adjusted odds ratios for mortality at the different centres grouped according to volume (figure 2) are similar and their confidence intervals overlap.

Mortality has been defended as an indicator of the quality of care in hospitals. In fact, mortality is an objective, reliable, precise, and bias-free measure that can be the direct consequence of substandard care; however, a high mortality rate does not always indicate poor quality and poor quality does not always result in greater hospital mortality [24]. In the United States, the Agency for Healthcare Research and Quality (AHRQ) has approved the use of hospital mortality rates for 8 surgical procedures as criteria of quality and possible referral of patients to other centres [25]. These 8 procedures were selected because of their high mortality and because of the high variability in mortality among the different hospitals that they analyzed. Nevertheless, as Dimick et al. point out, the low frequency of some of these 8 surgical procedures at some centres raises the question whether it is appropriate to use mortality rates as a measure of quality in all cases [5].

From the information available in our study, it is difficult to deduce what aspects of the process of care (details about surgical management, for instance) have led to complications such as peritonitis, kidney failure, or respiratory failure, and this makes it difficult to take action to improve the quality of care. Likewise, suture failure can occur after technically impeccable surgery, because it depends to a certain extent on other factors such as the patient's nutritional and/or immune status. This is one limitation of hospital mortality studies that use administrative databases if the aim is to use the results to improve the process of care.

Furthermore, as some authors have already noted, administrative databases also have limitations for adjusting patients' baseline risks to enable comparisons of mortality rates [26‐29]. These limitations are related to a) differences in (or the lack of) coding for some comorbidities or procedures and the consequent possibility of under-coding of diagnoses in patients with greater severity (for example, the variable relative to nutritional deficiencies), b) the misclassification of certain health problems, c) the failure to register some variables of known clinical importance (for example, the clinical stage of the tumour or indicators of the patient's pre-operative physiological state, which is of key importance in surgical patients [30, 31]), and d) the difficulty in distinguishing among health problems that were present before admission from those that might have resulted from complications of the healthcare process.

We used in-hospital mortality as the outcome variable in the study because it was the only mortality variable available in the administrative databases; however, using this outcome variable can lead to limitations in interpreting differences in mortality. For instance, in-hospital mortality would probably be higher in centers that prolong LOS than in centers with a policy to discharge patients earlier. Patients discharged earlier might die within 30 days of discharge and this would result in an underestimation of surgical mortality. Table 4 shows that LOS was lower in centers with a lower volume of interventions, and this could partially explain the lower in-hospital mortality in those centers. As mentioned above, the different studies published use different definitions of mortality; Table 6 shows the most recent results about volume of interventions and short-term mortality for gastric cancer.

According to our data, it is evident that the specific location of the tumour is also often under coded and that the mean number of secondary diagnoses registered varies among different hospital volumes. In our study, patients who died in hospital had a higher number of secondary diagnoses, and this confirms a certain register bias that favours patients who die in the hospital. On the other hand, the absence of tumour stage, a key factor for the patient's prognosis (especially long-term prognosis), is an evident limitation. We would expect only patients with the most advanced stages (although with the possibility of being cured by surgery, as in our study) to have a greater risk of in-hospital death after the intervention. Finally, some authors have advocated tumour resection with radical lymph-node excision (D2), claiming that long-term outcomes (survival) are better than with more conservative surgery (D1), although D2 also has greater post-operative morbidity and mortality that counteracts the possible benefits [32, 33]. The coding system used for the MBDS-HD, the ICD9CM, does not allow us to distinguish between these specific aspects of the care process, not only because of possible under coding, but also because of the lack of specific codes for this or other procedures.

Many of the limitations of this study derive from the fact that compiling homogeneous, highly reliable, and specific information is impossible due to the lack of information systems and specific clinical registers for oncological surgery in Spain, and this problem obviously needs to be tackled. Otherwise, it will be practically impossible to completely analyze the process of care and outcomes for oncological surgery that will enable us to take measures to improve the quality of care.

Even with the possible limitations, we have been able to determine that mortality is greater for tumours located in the fundus or even in the cardia; many tumours of the fundus also invade the cardia. Also, we have observed a lower crude mortality following a Billroth I than after a Billroth II in tumours located outside the cardia or fundus. In cases in which a Billroth I was used, we assume that it was a small tumour and that the patient could benefit, leading to low in-hospital mortality.

Administrative databases like the one used in this study are currently the only source of information that is common to all centres, homogeneous, accessible, and considerably exhaustive; they contain epidemiological and clinical information about the hospital discharges for a given diagnosis in the Spanish healthcare system and in those of many other countries [34]. Moreover, in-hospital mortality is a highly reliable objective measure that is available in all of these administrative databases and that can be monitored over time. Therefore, these databases can serve as the starting part for the analysis of the quality of care and detection of possible problems that might have an impact on hospital mortality.

We found no independent relation between adjusted hospital mortality and the volume of interventions at a hospital. Moreover, there does seem to be a similarity in the risk of in-hospital mortality and other indicators (LOS, Charlson score, urgent admission) among the centres that performed more than 17 interventions per year.

In a recent prospective study carried out in Scotland, Thompson et al. also found no relation between hospital volume and mortality after surgery for stomach cancer [35]. Two other studies carried out in the United Kingdom found a very weak relation favouring lower mortality with greater volume [23, 36]. In our study, the univariate analysis found a statistically significant association between volume and mortality in the opposite direction to that hypothesized, so that greater volume was associated to greater mortality. This tendency was only observed in tumours located in the cardia or fundus, but it was not significant and may be due to the low number of cases. This possible association was not significant in the multivariate analysis, either; again, this could be due to the low number of cardia or fundus tumours and the association found between the type of procedure and mortality for tumours outside those locations.

In fact, hospitals that treat a larger number of cases might care for patients with more severe disease and more comorbidities who have a higher probability of complications, and this might predispose to a higher mortality rate. In our study, we identified a significantly higher percentage of patients with Charlson score greater than or equal to 3 (see table 4) in higher volume centres, and a higher rate of complications of surgical procedures in those centres (data not shown).

However, some studies of mortality after surgery for stomach cancer have found that university hospitals and those that treat a greater volume of cases might have lower post-operative and long-term mortality rates, [10, 37‐41] although the results of other studies contradict this hypothesis or fail to confirm it [3, 4, 21, 35, 42‐44] and other authors have questioned this hypothesis with respect to surgery for other types of cancer [45].

Despite the limitations in the comparability of the studies, it seems clear that the results are contradictory and there is no consensus, as is shown in table 6. It is evident that the characteristics of the healthcare systems, such as patient referral practices, centralization of oncological surgery, and the financing of medical procedures varies widely between the United States, Japan, Canada, and most European countries in which these questions have been analyzed. In the Spanish National Healthcare System, all medical procedures are financed by the System and patients are free to choose the centre where they are treated. On the other hand, the most complex patients tend to be treated in the centres that have the most experience in this type of interventions, so there is a certain degree of centralization Differences in the degree of centralization might partially explain the disparity in the cut-off points used for the volume of interventions in the different studies. Greater centralization can also lead to two different effects. On the one hand, centres specializing in a certain intervention are likely to attract more complex or more severe patients with greater a priori possibilities to die in the short term; on the other hand, increased volume and greater experience are likely to counterbalance these effects on in-hospital mortality.

Thus, it seems logical that any interpretation of the results published should take into account not only the comparability of the design of the studies, but also the specific characteristics of each healthcare system and the time frame of the observations [46]. Furthermore, it might be a mistake to consider greater volume as a standard to predict better quality, when it is more likely the structures, the experience and specialization of the professionals, and the many different processes linked to this type of intervention that are responsible for better results, as many authors have pointed out [47‐49].