Figure 1. Attributable mortality due to the 25 most mortal individual complications listed by the International Classification of Diseases, Ninth Revision (ICD-9) code: 518.81, acute respiratory failure; 428.0, congestive heart failure, not otherwise specified; 427.31, atrial fibrillation; 518.5, pulmonary insufficiency following trauma and surgery; 348.8, brain conditions, not elsewhere classified; 507.0, pneumonitis due to inhalation of food or vomitus; 584.9, acute kidney failure, not otherwise specified; 348.4, compression of brain; 458.9, hypotension, not otherwise specified; 348.5, cerebral edema; 348.1, anoxic brain damage; 286.9, coagulation defects, not elsewhere classified and not otherwise specified; 403.91, hypertensive chronic kidney disease; unspecified; with chronic kidney disease stage V or end-stage renal disease; 427.41, ventricular fibrillation; 780.01, coma; 286.6, defibrination syndrome; 276.2, acidosis; 427.89, cardiac dysrhythmias, not elsewhere classified; 785.52, septic shock; 038.9, septicemia not otherwise specified; 410.91, acute myocardial infarction of unspecified site initial episode of care; 410.71, subendocardial infarction, initial episode of care; 276.7, hyperpotassemia; 434.91, cerebral artery occlusion not otherwise specified with cerebral infarction; and 253.5, diabetes insipidus.
Figure 2. Percentage of mortality attributable to complications by hospital (95% bootstrap CIs).
Customize your JAMA Network experience by selecting one or more topics from the list below.
Osler T, Glance LG, Hosmer DW. Complication-Associated Mortality Following TraumaA Population-Based Observational Study. Arch Surg. 2012;147(2):152–158. doi:10.1001/archsurg.2011.888
Context Complications are common in the care of trauma patients and contribute to morbidity, mortality, and cost. However, no comprehensive list of surgical complications is widely accepted.
Objectives To create an empirical list of complications based on the International Classification of Diseases, Ninth Revision (ICD-9) lexicon and estimate the contribution of these complications to mortality.
Design Retrospective database analysis.
Setting Office of Statewide Health Planning and Development data set.
Patients The Office of Statewide Health Planning and Development provided information on 409 393 patients admitted to 1 of 159 California hospitals between 2004 and 2008. We defined a complication to be any ICD-9– coded condition that accrued after hospital admission and significantly increased mortality.
Main Outcome Measures Odds of mortality for individual complications and number of excess deaths due to individual complications based on attributable risk fractions.
Results Eighty-two different ICD-9 codes contributed significantly to mortality as complications. Odds ratios ranged from 1.02 (hyperosmolarity) to 46.1 (ventricular fibrillation). There were a total of 175 299 complications (range, 0-14; average 0.4/patient). Twenty-four percent of patients had at least 1 complication. Mortality increased with the number of complications; each additional complication increased mortality by 8%. Absent any complications, there would have been 7292 fewer deaths, a 64% reduction in overall mortality.
Conclusions Most complication-related mortality is due to 25 individual complications. Eliminating all complications might prevent two-thirds of deaths, but because many complications are not preventable, this figure is the upper bound on possible mortality reduction. Hospitals vary in their proportions of deaths due to complications, and thus, efforts to prevent complications might improve survival at some hospitals.
Complications have always stalked the care of surgical patients in general and trauma patients in particular, but only in the last 2 centuries have surgeons come to regard complications as potentially avoidable.1 Because complications were both common and associated with increased morbidity, cost, and mortality, over time surgeons came to regard complications as a way to evaluate care and improve results. Despite this long relationship between complications and surgical care, there is surprisingly little consensus on which events should be regarded as complications2 and still less on how to measure the severity of individual complications. As a result, research on complications has depended on ad hoc definitions and measures of severity. Further, because discovery of complications often involves manual medical record abstraction, research data sets used in the study of complications have tended to be small, expensive, and hard to reproduce.
One alternative to small-scale intensive medical record review is the use of large administrative data sets. Although such data sets usually fail to capture which International Classification of Diseases, Ninth Revision (ICD-9) codes represent complications, we speculated that the “present on admission” (POA) indicator could be used to define a complication as any ICD-9 code that accrued after hospital admission and, further, that applying this empirical definition of a complication to a large administrative data set might allow us to objectively assess the severity of each type of complication and perhaps even answer the counterfactual question: “What would be the change in mortality rates if all complications could be eliminated?” Although in practice such a scenario is unlikely, this approach could provide an upper bound on how much improvement in overall mortality is possible by scrupulously preventing complications.
We chose to base our investigation of complications on a population of trauma patients for several reasons. First, because injury is a frequent cause of hospital admission, large numbers of trauma patients are available for study. Additionally, both complications and death are more common in the trauma population than in most other surgical conditions. Finally, models that predict mortality following injury have been well studied, and very recently, a trauma mortality model that relies solely on injury descriptions expressed in the ICD-9 lexicon has been described.3
The Office of Statewide Health Planning and Development collects data on all patients hospitalized in California and makes these data available to investigators. This data set fit our purposes well because it includes uniform information on a large population of trauma patients. Perhaps as importantly to this investigation, which depends heavily on the POA indicator, the State of California has collected the POA indicator as a required element in its uniform hospital discharge data set for more than 2 decades and thus has considerable experience with this data element. The Committee for the Protection of Human Subjects of the California Health and Human Services Agency approved this project.
The Office of Statewide Health Planning and Development provided information on 1 227 531 patients admitted to any hospital in California between 2004 and 2008 who had 1 or more ICD-9 codes within the general range of trauma (800-959.9). We excluded patients whose primary ICD-9 code was not in this range (398 355 exclusions as well as patients whose primary ICD-9 code suggested nonmechanical injury [burns, late effects of trauma, and foreign bodies; 65 094 exclusions]) or whose ICD-9 E codes suggested nonmechanical injury (eg, hanging, asphyxiation, and drowning; 25 590 exclusions). We excluded patients younger than 16 years (61 901 exclusions) and patients with isolated hip fractures older than 65 years (64 313 exclusions). At the hospital level, we excluded patients admitted to any hospital with fewer than 1000 trauma cases over the 5 years of our study (238 hospitals; 82 738 exclusions). We also excluded patients admitted to either of 2 hospitals with an unexpectedly low rate of POA assignation (7705 patients) or to the single hospital with an implausibly high rate of POA assignation (3647 patients). Finally, we excluded any patient missing any data required for our study: POA indicator, 22 067; age, 78; hospital identifier, 9; or vital status at discharge, 6567. Patients who were transferred to another hospital were considered to have an unknown vital status at discharge and excluded from our analysis. Our final data set consisted of 409 393 patients admitted for traumatic injury to any 1 of 159 hospitals over the 5 years from 2004 through 2008.
We first compared the overall rates of POA assignation among individual hospitals to assess the reliability of this data element and possibly exclude hospitals with either unusual coding practices or patient populations. We used the following 2-stage procedure for this comparison. We first fit a random-effects logistic model that predicted POA with age as the single level 1 predictor and hospital as the sole level 2 predictor. We then identified outlying hospitals by applying the Grubbs4 procedure to the theoretically normally distributed predicted values of the level 2 hospital effects.
We considered the 2786 different ICD-9 codes that were acquired subsequent to hospital admission at least once as candidate complications. Most of these codes were sparsely represented, however, and more than half (1781) were not associated with even a single death. We confined our search for complications to the 247 ICD-9 codes that were associated with 10 or more deaths in our data set because logistic models typically require 10 or more positive outcomes to reliably estimate an odds ratio for a predictor of interest. We entered these ICD-9 codes as 247 binary predictors into a single stepwise logistic mortality model that also controlled for age, traumatic shock, and extent of injury. (Extent of injury was included as a single term in this model, the logit transformation of the probability of death predicted by the Trauma Mortality Prediction Model [TMPM].)3 The odds ratios for each of the 82 ICD-9 codes that were significantly associated with mortality as complications (P < .01) were taken as an objective measure of that complication's severity in terms of its propensity to cause death. We also computed the attributable mortality fraction for each individual complication as well as for all complications in concert using the maximum likelihood estimation procedure described by Greenland and Drescher5 and implemented as the Stata punaf command.6 Finally, for each of the 159 hospitals in our data set, we computed the proportion of mortality attributable to complications along with bootstrap estimates of the uncertainty in these values. All data manipulation and analyses were carried out using Stata/MP (version 11.0; StataCorp).
The overall rates of POA assigned to ICD-9 codes of trauma patients varied among the hospitals, ranging from 86.2% to 99.9%. The claim by 1 hospital that 99.9% of all diagnoses were POA seemed unlikely to be true, and further, the range of values among hospitals suggested possible variability in coding practices, patient populations, or both. We excluded the 3 hospitals (10 722 patients) that fell outside the distribution of rates of POA of the other 159 hospitals in our data set as judged by the Grubbs4 procedure.
After exclusions, the hospitalizations of 409 393 trauma patients represented by a total of 2.5 million ICD-9 codes were available for study. Overall mortality was 2.8%. Most ICD-9 codes were coded as POA and thus represented either comorbidities (57%) or injuries (34.5%). The remaining 8.5% of ICD-9 codes were coded as occurring after hospitalization and thus possibly complications.
Our basic logistic risk adjustment mortality model controlled for age, traumatic shock on admission (defined as ICD-9 code 958.4 reported as POA), and overall extent of anatomic injury. This model displayed outstanding discrimination (receiver operating characteristic = 0.90). A calibration plot showed this model to be well calibrated throughout most of its range, although it slightly underpredicted mortality for the 2018patients (0.5%) with predicted mortality greater than 60%. The relatively large Hosmer-Lemeshow statistic (88.4) reflected the very large size of our data set, which allowed the Hosmer-Lemeshow test to detect a relatively small deviation from perfect calibration.
Eighty-two individual ICD-9 codes emerged from a stepwise logistic regression mortality model as complications that contributed significantly to mortality. The Table displays these diagnoses as well as their rates of occurrence, raw mortality, odds ratios of death, and attributable excess deaths.
Although 82 different ICD-9 codes are included in our list of complications, a few codes are virtually synonymous (eg, ICD-9 code 458.8 “hypotension, other specified” and ICD-9 code 458.9 “hypotension, unspecified”) as confirmed by their very similar odds of mortality ratios. On the other hand, some codes on our list that are often combined into a single complication category have very different individual mortality risks (eg, 4 different ICD-9 codes describing cardiac arrhythmias display a range of odds ratios, from 1.41 [cardiac dysrhythmia] to 46.1 [ventricular fibrillation]).
On average, each patient experienced 0.4 complications. Twenty-four percent of patients experienced 1 or more complications, with most patients experiencing a single complication. The risk of mortality increased almost linearly with number of complications, with an attendant increase in mortality of approximately 8% for each additional complication.
The attributable risk fraction due to all complications in aggregate was 0.64. As expected, attributable risk at the level of individual complications varied widely. A Pareto graph suggests that most deaths were due to a handful of complications (Figure 1). A single complication, respiratory arrest (ICD-9 code 518.81), accounted for 9% of attributable deaths. A group of only 8 complications were responsible for 50% of the attributable deaths, and 25 complications contributed 80%. The average attributable risk at the hospital level for all complications in aggregate was 54% but varied from a low of 37% to a high of 65% (Figure 2). No individual hospital stood out in this analysis, although 25 hospitals had values that differed from the mean value at a statistically significant level (P = .05).
The concept of a complication has long played a central role in how clinicians conceive and understand the course of medical care, and thus, investigators have naturally tried to use this construct in the evaluation of surgical care. Such efforts have usually involved a predetermined list of possible complications that are assigned to patients either by medical record review7,8 or by mapping previously assigned ICD-9 codes onto the list of predetermined complications.9,10 Our approach differs fundamentally from these efforts by reversing the paradigm and allowing individual complications to emerge from a data set based solely on their statistical association with mortality.
By allowing the data to speak for itself, we arrive at a list of complications that will seem grimly familiar to clinicians. Moreover, the odds ratios may have face validity as a relative measure of complication severity. Our list of complications consists almost entirely of acute physiologic events, including primary cardiac (arrhythmias, myocardial infarction, and heart failure), renal (oliguria, anuria, and renal failure), pulmonary (respiratory failure, acute respiratory distress syndrome, pulmonary emboli, and pneumothorax), central nervous system (convulsions, encephalopathy, coma, edema, cerebrovascular accident, and anoxic brain damage), and infectious (pneumonia, sepsis, and septic shock) events. A few complications are events with typically slower onset that occur later in the course of hospitalization (gangrene and intestinal obstruction). We intentionally excluded a single ICD-9 code, cardiac arrest (427.5), because this code could be legitimately coded for every patient who died.
Notable for their absence from our list are several classic surgical complications such as wound infection. We believe this omission is appropriate, because while wound infections certainly add to morbidity, cost, and length of stay, they likely contribute only slightly to mortality. Indeed, many traditional surgical complications (eg, abdominal abscess following laparotomy) that formerly presented some risk of death but today are readily treated are absent from our list. More generally, our list emphasizes complications that are both intrinsically difficult to prevent and carry a substantial risk of morality. Many events commonly identified as complications by clinicians (eg, pulmonary embolus, acute respiratory distress syndrome, and food/vomit pneumonitis) appear on our list because, despite the prophylactic measures that are currently available, these events remain common and present significant risk of death. A list of complications based on prolongation of hospital stay rather than mortality might be quite different from the list we present herein because it would likely consist of untoward medical events more amenable to treatment. Such a length of stay–based examination of complications will be the subject of a future report.
Our list of 82 complications reflects some of the idiosyncrasies of the ICD-9 lexicon. A few complications seem to be synonyms that could be combined in the interest of parsimony. On the other hand, some ICD-9 codes appropriately distribute some summary diagnoses (eg, arrhythmia) into several different complication diagnoses with rather different odds ratios, ranging from “atrial fibrillation” with an odds ratio of 1.491 to “ventricular fibrillation” with an odds ratio of 46.1. In this case, combining different types of arrhythmia into a single complication would overlook an important distinction.
We calculate that the fraction of deaths attributable in aggregate to the complications on our list is 64%, a result that implies that 64% of deaths could be avoided if all complications could be eradicated.11 However, because not all complications can be prevented by better use of available therapies, this figure is an upper bound on the number of avoidable deaths. Indeed, close examination of our list of complications discloses some complications for which there are no reliable prophylactic measures. A Pareto diagram of attributable deaths by complication suggests, for example, that 9% of deaths could be avoided if respiratory failure were eradicated. However, clinicians might feel that respiratory failure is in many cases a marker for a patient who is simply too weak to breathe without assistance rather than a specific, preventable clinical event. Respiratory failure in such cases represents the culmination of a host of other, perhaps preventable, events that have exhausted a patient's reserves. Other complications on our list represent specific, possibly preventable clinical events (eg, pulmonary embolus) that perhaps comport more closely with clinicians' idea of a complication. However, when such complications are relatively rare, even if associated with a substantial risk of mortality, they contribute few attributable deaths overall; pulmonary emboli were responsible for only 10 attributable deaths per year despite an odds ratio of 3. Conversely, more common complications that may not be strongly associated with death, such as pneumonia (4031 cases/5 y; odds ratio of 1.36), also contribute only modestly to attributable deaths, again adding only about 10 deaths per year (Figure 2).
Complications are of perennial interest to surgeons and have been the subject of many investigations. A recent report from Ingraham et al,9 who examined almost 100 000 patients provided by the National Trauma Data Bank, confirms our observation that cardiovascular events, acute respiratory distress syndrome, renal failure, and sepsis are significantly associated with death but fails to confirm our finding that pulmonary emboli and pneumonia are also causes of increased mortality, perhaps because its case-control design provided less power than our regression approach. We concur with Ingraham et al that surgical infections, urinary tract infections, and Clostridium difficile infections do not increase mortality. Overall, our point estimates for attributable mortality tend to be somewhat higher for individual complications than those found by Ingraham et al.
We believe we have derived the first empirical list of complications for trauma patients. One use for this list might be to identify complications that are linked to the greatest number of deaths (eg, atrial fibrillation). Such a list could suggest specific complications that clinicians might wish to target, perhaps with specific clinical protocols, to avert the greatest number of fatalities. Conversely, complications on our list that are associated with low mortality burdens (eg, cerebral embolism with infarct) might be given less priority.
Our finding that hospitals differ in their rates of mortality attributable to complications might suggest that our list of complications be used as a basis for comparing hos pitals' quality of care. However, because hospitals are likely to differ in their distribution of patients at risk for various complications, comparing hospitals on this metric will require adjustments for patient mix that in practice might prove too complex to be useful. It might be possible to compare hospitals on the rates of individual complications that are common and associated with substantial mortality (eg, acute kidney failure) or are simply associated with very high risk of mortality (eg, anterior myocardial infarction), but individual risk adjustment models would be required for each complication of interest. Such models would be difficult to develop for rare complications such as hepatorenal syndrome.
Our study has several potential limitations. Although our data set is large, it is likely that some complications are so rare or so rarely associated with death that we were unable to detect them. However, because any complications missing from our list are rare or rarely fatal, we believe that our list of complications may prove useful to physicians, researchers, and administrators despite its being incomplete. As with all administrative data sets, coding of some complications may have been influenced by the outcome of interest (death). Moreover, some instances of complications certainly went unrecorded by hospital coders. In earlier studies of administrative data sets, sensitivities as low as 0.5 for some complications have been observed. We believe that our data set may have more faithfully recorded complications because we were careful to exclude hospitals with implausibly low rates of POA assignation and because the 82 ICD-9 codes that define our list of complications are dramatic events that may be less likely to be overlooked than other events that are less closely tied to mortality. Importantly, while any underreporting of complications will obviously affect our estimates of both the prevalence and attributable mortality of individual complications, it is unlikely to greatly change our estimate of their odds of mortality.
Complications have been used as a tool in the evaluation of surgical care for almost a century but this approach has been troubled by a lack of consensus on which events should be considered complications. Large administrative databases are now available that distinguish between diagnoses that are POA and those that develop subsequently and we find that this distinction can be used to empirically define ICD-9– coded events as complications and, further, to compute both the severity and the mortality burden of each complication so defined. We believe that the empirical study of complications in large populations may help illuminate the role of complications in medical outcomes. However, because some complications are relatively rare and many complications may not be preventable, the use of complications to compare the quality of care among hospitals is likely to remain challenging.
Correspondence: Turner Osler, MD, MSc, Department of Surgery, University of Vermont, 789 Orchard Shore Rd, Colchester, VT 05446 (email@example.com).
Accepted for Publication: August 25, 2011.
Author Contributions: Dr Osler had full access to all the data in the study and takes responsibility for the integrity of the data and the accuracy of the data analysis. Study concept and design: Osler and Glance. Acquisition of data: Osler. Analysis and interpretation of data: Osler, Glance, and Hosmer. Drafting of the manuscript: Osler and Hosmer. Critical revision of the manuscript for important intellectual content: Osler and Glance. Statistical expertise: Osler, Glance, and Hosmer. Obtaining funding: Glance. Study supervision: Osler.
Financial Disclosure: None reported.
Funding/Support: This work was supported by grant RO1 HS 016737 from the Agency for Healthcare Research and Quality. The Office of Statewide Health Planning and Development data set was provided by the State of California Office of Statewide Health Planning and Development.
Disclaimer: The opinions expressed herein are those of the authors.