Postoperative Mortality in The Netherlands: A Population-based Analysis of Surgery-specific Risk in Adults

• ❖ Perioperative mortality in large, unselected populations of surgical patients has not been accurately assessed

• ❖ In a review of 3.7 million surgical procedures at 102 hospitals in the Netherlands during 1991–2005, perioperative mortality was 1.85%

• ❖ Mortality differed by more than 250-fold among surgical procedures, and a model based on 36 surgical subcategories predicted mortality better than a simple low versus  high surgical risk model.

MANY patients consider surgery as a major life event because it disrupts their personal, professional, and economic life, as well as their physical body.1Fear of being anesthetized and undergoing surgery are commonly mentioned causes of anxiety.2During preoperative evaluation, a patient is informed of significant risks of anesthesia and surgery, which is essential in receiving fully informed consent. In this respect, information on surgery-specific mortality rates is crucial.

Estimates of perioperative mortality in patients undergoing anesthesia vary largely and range from 1 death in 53 anesthetics to 1 in 5,417.3,4Reports on the incidence of perioperative mortality that is directly attributable to anesthesia show values ranging from 1 in 6,795 to 1 in 200,200 anesthetics.3It is suggested that these wide ranges may be caused by differences in operational definitions and reporting sources.3However, the diversity of the studied surgical cohorts is a—if not the—major contributor to the observed variability in patient outcome too.5–8Currently, few data are available that systematically describe rates and trends of postoperative mortality for fairly large, unselected patient populations. This study was undertaken to overcome this lacuna. We present estimates of the surgery-specific postoperative mortality risk for all commonly performed procedures based on an analysis of 3.7 million surgeries that were conducted in The Netherlands between 1991 and 2005.

Data were obtained from Prismant.#Prismant itself obtains information from the National Medical Registry (Dutch: “Landelijke Medische Registratie”[LMR]). LMR was launched in 1963 to systematically collect medical and administrative data on each admission to a general or an academic (teaching) hospital in The Netherlands. Since 1986, LMR provides nationwide coverage as data are delivered by any hospital in The Netherlands, except by the Antoni van Leeuwenhoek Cancer Institute in Amsterdam. On January 1, 2007, there were 102 participating hospitals, which delivered information on 3.3 million admissions in that year, including 1.5 million admissions for surgical or diagnostic procedures.

The information that is collected by LMR includes the main (discharge) diagnosis, secondary diagnoses, dates of admission and discharge, death during admission, operations and related therapeutic procedures, age, sex, and comorbidities that are classified according to the International Classification of Diseases 9th revision Clinical Modification (ICD-9-CM).**LMR contains data on inpatient and outpatient procedures, open and laparoscopic procedures, day-case surgery, and surgery requiring hospitalization. Although the LMR database is not designed for scientific applications, but for administrative purposes, the information system can be considered an adequate resource for epidemiologic studies. Surgeons have an incentive to provide numbers and data that are correct and up to date because their information is needed to maintain their national medical license to operate (the data that they provide is subject of inspection by independent national health care inspectors). Hospital administrators have an incentive to adequately report data because the LMR registry forms the base of reimbursements of surgery-related costs by health insurance companies. Recently, the LMR has been validated, and data items that appeared in the LMR database were compared with source documents (medical charts) in 55 hospitals. There was 99% agreement between the two files with respect to general administrative data and 92% agreement with respect to data on surgical procedures.9,10 

This study included all elective, nonday case, open surgical procedures that were performed in The Netherlands in the 15-yr period, from January 1, 1991, to December 31, 2005, in patients aged 20 yr or older. Thus, procedures that were performed laparoscopically, as well as (diagnostic) procedures and interventions that are typically performed by radiologists or gastroenterologists, were excluded. We also excluded surgeries that were considered “urgent” or “emergent.” Individual data on date of birth, sex, date of admission, date of surgery, type of surgery, date of discharge, hospital mortality, as well as on the following comorbidities: diabetes mellitus (ICD-9-CM 250), renal insufficiency (ICD-9-CM 580), ischemic heart disease (ICD-9-CM 410, 411, 412, 413, and 414), heart failure (ICD-9-CM 428 and 429), cerebrovascular disease (ICD-9-CM 430), cardiac arrhythmia (ICD-9-CM 426 and 427), hypertension (ICD-9-CM 401, 402, 403, 404, and 405), peripheral arterial disease (ICD-9-CM 440, 441, 442, 443, 444, 445, 446, 447, and 448), and pulmonary disease (ICD-9-CM 490, 491, 492, 493. 494, 495, and 496) were obtained.11Data were transferred from Prismant to the Erasmus MC and entered in an electronic database that was prepared for statistical analyses.

The unit of observation in LMR is the hospital admission, and patients are included as many times as they had a new admission. For privacy reasons, Dutch regulations do not allow LMR to contain unique patient identifiers. Therefore, we had to select the surgical procedure and not the patient as unit of analysis. In fact, this is consistent with clinical practice because the risk of perioperative adverse outcome is usually assessed in relation to a specific procedure. Furthermore, this approach guarantees an optimal utilization of the available information. Our dataset may contain reoperations or operations for complications of the initial procedure. However, as a consequence of the LMR system, this will only be the case in patients who had initially been discharged (alive).

In the Netherlands, surgical procedures are classified at admission by the treating physician according to a standardized national classification system (Dutch: “Classificatie van Verrichtingen”).12These data are stored in the LMR database. For the purpose of our study, we grouped surgical procedures, using the national classification system, into 11 main categories (abdominal, orthopedic, urologic, breast, neurologic, gynecology, vascular, cardiac, endocrinology, pulmonology, and ear, nose, and throat) and 36 subcategories according to generally accepted medical theory and practice (these categories are listed in the Results section). Grouping of surgical procedures was performed independent of the statistical analyses.

In the LMR, the vital status was not documented for each patient at the same fixed time point after surgery. Instead, the LMR contained information on vital status until the day of hospital discharge, which varied from patient to patient. Therefore, we have chosen “postoperative” all-cause death as the primary endpoint, which was defined as death from any cause occurring during the initial hospital stay after surgery or within 30 days after surgery, whichever date came first.

Continuous data are described as median values and corresponding 25th and 75th percentiles, and dichotomous data are described as numbers and percentages. Differences in baseline characteristics according to type of surgery were evaluated. However, we did not perform formal statistical testing to draw inferences from the sample because even small, clinically irrelevant differences will likely be statistically significant because of the large number of patients being analyzed, whereas relevant differences will be visible anyhow.

Univariable and multivariable logistic regression analyses were applied to evaluate the relationship between type of surgery and the primary endpoint. The logistic regression model requires a reference category for each of the included explanatory variables. Because there is no surgical procedure that can be considered as a “natural” reference, we determined the log odds (i.e ., the “logit”) of the probability of the primary endpoint for each procedure relative to the (unweighed) average logit over all procedures. This method is known as the “deviation from means” approach.13We present crude, unadjusted odds ratios (OR) and their 95% confidence intervals, as well as ORs that are adjusted for (potential) confounding factors, including age, sex, and the comorbidities that are listed under “selection of patients and procedures” in the Materials and Methods section.

We did not intend to develop a mortality risk prediction model. However, to better understand the relation between type of surgery and postoperative outcome, we studied the performance of the multivariate models with respect to discrimination. Discrimination in the context of this study refers to the ability of the model to distinguish between patients with and without postoperative death. It was quantified by the c-statistic, which may take any value between 0.5 and 1 and indicates better discrimination if closer to one.14 

All analyses were undertaken in SAS (SAS Institute, Inc., The SAS System for Windows version 8.02. Cary, NC).

In the literature, several risk indices are described that relate patient characteristics and type of surgery to a variety of adverse outcomes. The Revised Cardiac Risk Index by Lee et al .7is developed for the prediction of perioperative cardiac complications. Surgery is classified as high risk, including intraperitoneal, intrathoracic, and suprainguinal vascular procedures, or nonhigh risk. The index that is based on the National Veterans Affairs Surgical Risk Study aims to predict the risk of all-cause perioperative death.15A total of six surgical categories are distinguished, including intrathoracic surgery (OR 4.4, relative to breast, urologic, and endocrine surgery, the categories with lowest mortality incidence in that study), neurosurgery (OR, 2.7), general surgery (OR, 2.6), otolaryngology (OR, 2.3), peripheral vascular surgery (OR, 2.2), and orthopedic surgery (OR, 1.6). The Physiologic and Operative Severity Score for the enUmeration of Mortality and Morbidity index is developed to estimate the risk of all-cause mortality and a wide variety of morbidity conditions, ranging from wound infection to cardiac failure.16The Physiologic and Operative Severity Score for the enUmeration of Mortality and Morbidity index separates four surgical categories with an “operation severity” that is described as minor, moderate, major, and complex major.

To evaluate the clinical usefulness of dividing types of surgery into 36 subcategories (see Results section) relative to other risk indices, we applied the Revised Cardiac Risk Index, the index based on the National Veterans Affairs Surgical Risk Study, and the Physiologic and Operative Severity Score for the enUmeration of Mortality and Morbidity index to our database and evaluated their discriminative properties for all-cause mortality.

A total of 3,667,875 surgical procedures were included in this analysis. Abdominal (33% of the study cohort), orthopedic (15%), and urologic (12%) procedures were most common (tables 1 and 2). The entire cohort consisted of slightly fewer men (48.1%) than women and had a median age of 62 yr. There were relevant differences in these demographics between the distinguished procedures. Naturally, breast and gynecology surgery were (almost) exclusively performed on women, whereas men underwent the majority of urologic (prostate) procedures. In addition, men more commonly had abdominal (hernia), vascular, cardiac (coronary artery bypass grafting; heart transplant), pulmonology, or ear, nose, and throat surgery. Women more commonly underwent orthopedic (hip) or endocrinology (thyroid) surgery. Patients who underwent orthopedic (hip), urology, or vascular surgery had highest median age and constituted an elderly cohort. Patients who underwent breast, neurologic, or endocrinology surgery had the lowest median age.

The prevalence of comorbidities varied largely between the surgical procedures (tables 1 and 2). In general, a high prevalence of comorbid conditions was observed in patients undergoing vascular or cardiac procedures. More specifically, a high prevalence of diabetes mellitus was observed in patients undergoing biliary duct or pancreatic surgery, and a high prevalence of diabetes mellitus, hypertension, or renal disease was seen in those undergoing organ transplants. Patients undergoing appendectomy, knee, breast, or gynecology surgery had low a prevalence of comorbidities.

The surgical population underwent some changes during the 15-yr study period (fig. 1). The strongest trend was observed in the number of orthopedic procedures, which increased from 7,500 (men) and 18,500 (women) in 1990 to 15,200 (men) and 32,000 (women) in 2005, implying an average annual increase of 5–6%. An interesting trend was observed in the number of pulmonology procedures as well, which steadily decreased in men and increased in women.

Along with these changes, clear time trends were also observed with respect to age, sex ratio, and prevalent pulmonary disease (fig. 2). Between 1991 and 2005, the median (25th–75th percentile) age increased from 60 (44–71) to 63 (50–74) yr, whereas the percentage of men decreased from 48.6 to 47.2%. The prevalence of chronic obstructive pulmonary disease decreased from 2.0 to 1.0%.

The endpoint of postoperative all-cause death was observed in 67,879 patients, who represented 1.85% of the study cohort. There were major differences in mortality in relation to type of surgery (table 3). Breast surgery was associated with lowest mortality (incidence, 0.07%; unadjusted OR, 0.07), and vascular surgery was associated with highest mortality (incidence, 5.97%; unadjusted OR, 6.56). Men undergoing orthopedic or pulmonology surgery had higher mortality than women (fig. 3), and men undergoing abdominal, urology, or cardiac surgery had lower mortality than women. We did not observe apparent trends in postoperative mortality over time (fig. 3), implying that time itself was not a major outcome determinant, so that we were allowed to merge annual datasets for further analyses.

After adjustment for age, sex, and comorbidities, breast (adjusted OR, 0.13) and gynecology (0.27) surgery were associated with a more than twofold lower than average mortality risk, whereas pulmonology (adjusted OR, 3.79), abdominal (3.03), and neurologic surgery (2.64) were associated with a more than twofold higher than average risk (table 3; fig. 4). The relative mortality risk that was associated with the other main surgical categories was within the range twofold lower to twofold higher than average.

Analyses of mortality figures according to 36 surgical subcategories are presented in table 4and figure 4. Patients undergoing liver transplant had almost 16-fold increased mortality risk relative to the average (adjusted OR, 15.78), whereas patients undergoing herniated nucleus pulposus surgery had more than 16-fold reduced risk (adjusted OR, 0.05). Procedures with an 8- to 16-fold increased risk relative to the average included lung transplant and cardiac congenital surgery, and procedures with a four- to eightfold increased risk included spleen, liver, gastric, and pancreatic surgery, as well as brain surgery and heart transplant. Procedures with an 8- to 16-fold decreased risk relative to the average included knee and breast surgery, and procedures with a four- to eightfold decreased risk included prostate, hernia, gynecology, myelum, and thyroid surgery.

Direct comparisons of absolute mortality risks that are associated with different types of surgery are usually hampered by confounding factors, including age, sex, and the patient's general condition or by differential patient selection. Indeed, in our dataset, demographic characteristics and comorbidities were associated with type of surgery (tables 1 and 2) and postoperative mortality (tables 3 and 4), and adjustment for these factors resulted in better discrimination (table 5). Figure 4presents estimated absolute mortality risks in relation to type of surgery for the “average” patient in the study cohort (i.e ., the virtual patient with characteristics equal to the values that are presented in the bottom rows of tables 1 and 2). According to the most extensive model that is based on 36 surgical categories, expected mortality ranged from 0.7 per 1,000 average patients undergoing herniated nucleus pulposus surgery to 185 per 1,000 average patients undergoing liver transplant.

The c-index of the multivariable model that included the binary surgical classification of the Revised Cardiac Risk Index was 0.829, similar to the value that was obtained with a model in which the surgical classification system of the National Veterans Affairs Surgical Risk Study was applied (table 5). With a c-index of 0.873, the model that included the classification of Physiologic and Operative Severity Score for the enUmeration of Mortality and Morbidity index had better discriminative power. In an ad hoc analysis, based on the results that are presented in figure 4, we separated five categories of surgery-related mortality risk: surgeries with an (adjusted) mortality < 2⁻², ≥ 2⁻²to 2⁻¹, ≥ 2⁻¹to 2, ≥ 2 to 2²and ≥ 2⁻²times the average mortality risk. The model that was based on this classification had a c-index of 0.875, virtually similar to the most extensive model that was based on 36 surgical subcategories.

According to the model in which the Revised Cardiac Risk Index was applied, 90% of the patients who experienced all-cause death had an estimated mortality of more than or equal to 1%. Hence, the sensitivity of the model at the threshold of 1% was 90% (fig. 5). The percentage of patients with an estimated mortality of less than 1% among those who stayed alive was 58% (i.e ., the specificity at the threshold was 58%). At the same 90% sensitivity threshold, the National Veterans Affairs index had a similar specificity level (57%), whereas the Physiologic and Operative Severity Score for the enUmeration of Mortality and Morbidity index, and our model based on 36 surgical subcategories had an apparently higher specificity: 69% for both models.

This observational, population-based study, involving 3.7 million surgical procedures in The Netherlands, provided a detailed and contemporary overview of postoperative mortality for the entire surgical spectrum. As such, it may act as a reference standard for surgical outcome in Western populations. In general, undergoing surgery can be considered a safe procedure because 70% of the procedures were associated with a mortality risk of less than 1%. Nevertheless, a large gradient in postoperative mortality was revealed in relation to type of surgery. After adjustment for differences in distribution of demographic characteristics and comorbidities, there seemed to be a more than 256-fold difference in the incidence of all-cause death between the highest and lowest risk categories. Hence, for purposes of clinical risk stratification, a simple classification of surgical procedures as high or low risk seems inappropriate.

The predicted risk of postoperative complications in individual patients is currently primarily based on the presence of comorbidities and not on the surgical procedure itself. Indeed, the severity of the underlying disease and the urgency of the procedure are important elements in the clinical decision-making process. However, physicians tend to balance the potential benefits of a procedure to the physical condition of the patient at risk, which is commonly defined by indexes using markers of cardiac, respiratory, and neurologic disease, with additional laboratory and electrocardiographic results. These risk factors are subsequently introduced in risk indices, which are useful to translate the patient's unique profile into suspected postoperative outcome. Our study suggests that predictions of postoperative outcome will improve if the severity of the surgical procedure will be given a higher weight.

Every surgical procedure elicits a stress response, initiated by direct tissue injury, change in body core temperature, pain, and anxiety. This response induces a systemic stress due to changes of sympathetic and parasympathetic tone and catecholamine surge, resulting in tachycardia and hypertension, further aggravated by perioperative fluid shifts. Surgery will also induce hypercoagulability caused by changes in the balance between prothrombotic and fibrinolytic activity. The surgery-induced stress may finally result in myocardial ischemia (or infarction), pulmonary insufficiency, and hyperglycemia, which are key determinants of postoperative survival, whereas the severity of the stress response is directly related to the extent and duration of the surgical procedure. The Prismant dataset does not contain detailed information on the magnitude, duration, location, blood loss, and fluid shifts related to the specific procedure, nor does it hold data on the cause of death—and this is a limitation of our analyses. However, our data are in agreement with observations that extensive and lengthy procedures are associated with an increased risk of adverse outcome.

One might argue that a rough estimation of surgery-related risk will be sufficient to decide on additional preoperative diagnostics and the perioperative installation of protective pharmacotherapy. For example, the American College of Cardiology/American Heart Association guidelines on preoperative cardiac risk assessment consider three surgical categories, with an estimated “low” (< 1%), “intermediate” (1–5%), or “high” (> 5%) incidence of cardiac events, and treatment recommendations are based on this classification.17Indeed, our data indicate that the improved diagnostic performance of risk indices that are based on more detailed surgical classifications is particularly relevant for patients at low (< 1%) or high (> 10%) mortality risk. However, it is relevant, not only for the patients in the low- or high-risk ends of the spectrum but also for benchmarking and quality-of-care purposes, to know whether the suspected probability of death associated with a certain surgical subtype is 1 of 100 (appendectomy) or 1 of 10,000 (herniated nucleus pulposus surgery; both procedures are considered low risk), or 1 of 20 (gastric surgery) or 1 of 7 (liver transplant; both procedures are considered high risk).

We recognize that our work has several limitations. First, data were derived from a database that was designed for administrative and not for scientific purposes. The administrative personnel who are responsible for LMR data management retrieve the information from patient dossiers, which are most often paper based. In practice, a suboptimal match may appear between the definition of a certain comorbidity according to the ICD-9 system and the information that is collected (by healthcare professionals) in the dossier. The data management personnel are instructed to be restrictive in such situations and avoid overreporting rather than underreporting. As a result, medical conditions might have been overlooked, and consequently, the relative contribution of these factors to postoperative death might have been underestimated.18 

Second, we were only able to adjust the relation between type of surgery and all-cause death for a limited number of comorbidities that were coded according to the ICD-9 system. Therefore, residual confounding may still exist. Additional (potential) confounding factors that were not available in the LMR database included the indication for the procedure, the clinical condition of the patient before the procedure, and pharmacotherapy.

Third, as per design, analyses were restricted to patients who underwent surgery. No information was included from patients who were screened but did not undergo surgery because their mortality risk was perceived as prohibitive. Obviously, exclusion of patients at risk of adverse outcomes might have diluted estimates of relative risk.

Finally, we only included elective, nonday case, open surgical procedures. One might argue that this choice has resulted in a falsely high estimate of the death rate for procedures that also could have been done laparoscopically (with presumably a lower complication rate than open procedures) or, oppositely, in a falsely low estimate of the death rate if we think sicker patients may be referred to interventionalists. However, taking our selection for granted, the presented surgery-related death rates are neither systematically too low or too high. For example, the adjusted incidence of all-cause death after open appendectomy was 0.74%, and this is a valid estimate of the mortality risk associated with this open procedure. However, from the perspective of the disease that necessitates an appendectomy, this percentage might be an over- or underestimation, depending on the applied surgical technique (open or laparoscopic) and the patient's clinical condition.

This population-based study provided a detailed and contemporary overview of postoperative mortality for the entire surgical spectrum, which may act as a reference standard for surgical outcome in Western populations. These data might be relevant for patients, to provide truly informed consent, as well as their caregivers, who should be interested to benchmark their personal surgical performance.