Timing of surgery for hip fracture and in-hospital mortality: a retrospective population-based cohort study in the Spanish National Health System

A retrospective cohort of patients of 60-years-old and over hospitalized for hip fracture during 2002 to 2005 in all public hospitals in 8 Spanish regions. Hospital discharge administrative databases were used to follow up patients from admission to discharge and identify the time to surgical repair and in-hospital mortality.

Spain is divided into 17 autonomous regions known as "autonomous communities" with a high degree of self-government, including responsibility for health care. Each Spanish regional government operates an extensive network of hospital and primary healthcare centres that provides free care to about 97% of its respective populations [27]. We use data from 8 autonomous communities (Andalusia, Aragon, Asturias, Basque Country, Valencia Community, Navarre, Galicia and Extremadura) participating in the Spanish Atlas of Medical Practice Variations project [28], a research project emulating the Dartmouth Atlas of Healthcare [29]. Other autonomous communities were not included because their data were not available at the start of the study or had quality problems in some of the study's key variables. The autonomous communities included had around 21.5 million inhabitants in 2005, approximately half of the Spanish population that year.

The primary source was the so-called "Conjunto Mínimo de Datos Básicos" (Minimum Basic Data Set, MBDS) of hospital discharges produced between 2002 and 2005 in the public hospitals included in the study. The MBDS, a homogeneous register used by all sNHS hospitals, systematically records administrative and clinical data from each hospital discharge with, among other, data about diagnoses and procedures (using the International Classification of Diseases 9th Revision Clinical Modification, ICD9CM), dates of admission, discharge and surgery, and reason for discharge.

From the 81,740 discharges with a main diagnosis of hip fracture (ICD9CM: 820.xx) produced between January 2002 and December 2005 in the 131 public hospitals in the 8 participating regions, and to ensure a more homogeneous premorbid health status between groups, we excluded 6,889 (8.43%) patients because they were younger than 60-years-old and/or had a diagnosis of multiple or pathological fractures (presumably non-osteoporotic) and/or were elective admissions (suggesting inpatient fractures and that the time from admission to surgery could be not related to the hip surgery) and/or had a length-of-stay longer than 30 days (delays beyond 30 days for non-medical reasons were very unlikely and clinical delays could correspond to very severe patients whose risk differences cannot be adequately adjusted with the available covariates).

Additionally, we excluded all patients from 19 hospitals (n = 74) with less than 30 hip fracture discharges in the period studied (because we suspect that these hospitals do not have traumatology and orthopaedic wards and cases correspond to patients transferred from other hospitals), 3 hospitals (n = 408 cases) with no deaths registered and 16 hospitals (n = 9760) with more than 10% of discharges without the surgical data recorded (to exclude hospitals with poor quality MBDS registers). From the remaining cases, we excluded patients without the surgical treatment approach (n = 7.273; 11,26%) or without a surgery date (n = 836; 1,02%). The final sample (see Figure 1) included 56,500 patients of 60 years old and over admitted with a hip fracture as a main diagnosis through the emergency department in a public hospital and receiving surgical repair during the episode of care.

In-hospital mortality identified in the discharge criteria codes of the Hospital Discharge MBDS as death or an "in extremis" discharge.

Time to surgery measured as the difference (in days) between the admission date and the surgery date. Because the MBDS does not record the time of admission we were not able to define 24 or 48 hour periods and the exposure variable was grouped as early surgery (repairs that were carried out on the same day or the day following admission) and delayed surgery (repairs that were carried out from the 2nd to the 30th day following admission).

Variables included in the study that might be associated with the use of early surgery as well as with the probability of death were: age (divided into four categories: 60-69, 70-79, 80-89 and 90-years-old and over), sex, main diagnosis (grouped into three groups: intracapsular fractures (ICD9CM: 820.0x a 820.1x), trochanteric fractures (ICD9CM: 820.2x a 820.3x) and non-otherwise specified fractures (ICD9CM: 820.8 y 820.9), number of secondary diagnoses (grouped into four categories: 0-1, 2-3, 4-6, and > 6), surgical procedure (distinguishing between fixation and arthroplasty), Risk Mortality Index (RMI) modified from the one used by Majumdar et al [15, 30] (unfortunately MBDS does not allow differentiation between comorbidities and complications because it does not record the onset of a diagnosis of complications; see Additional File 1 for codes and weights used), Charlson Comorbidity Index calculated following an adaptation of the Charlson Index [31] for administrative databases [32, 33]. Our version, as in other Charlson Index adaptations for administrative databases [34, 35], distinguishes between codes defining chronic conditions (always accounted for) and acute conditions (those not to be considered in the current admission but in the following admissions; see Additional File 2 for codes and weights used). Finally, the year of discharge (from 2002 to 2005) and whether the admission occurred on weekend or weekdays were also studied.

The databases used in this study were transferred to the researchers without any personal identification. The study, observational and with retrospective non-identifiable data, did not require specific approval by the Ethics and Clinical Trials Committees, or informed consent.

First, we described the characteristics of patients and surgery and a bivariate analysis was carried out (using odds ratios with their respective confidence intervals of 95%, CI95%) to explore the possible associations between these factors and the performance of early-surgery and in-hospital mortality. Second, we used a backward-forward multivariate logistic regression model to analyze the relationship between selected explanatory factors (age, sex, type of fracture, type of surgery, Charlson Comorbidities Index scores, RMI scores and discharge year) and the realization of early surgery. Third, a second backward-forward logistic regression model with mortality as a dependent variable was carried out. We used the same explanatory variables as in the previous model and also our main exposure variable: the timing of surgery. The performance of the models was evaluated using the C-Statistics for discrimination and the Hoshmer-Lemeshow test for calibration.

Additionally, we carried out some complementary analyses: 1) to explore the relationship between the proportion of delayed patients and the in-hospital mortality rate at the hospital level (see Additional File 3), 2) we use multilevel models (a random intercept model adding a random-slope onto early surgery) to explore the impact of the differences between hospitals in the effect of early surgery on mortality (see Additional File 4), and 3) we use an instrumental variable approach (using the day-of-the-week admission as an instrumental variable) to validate the main findings of the study [36] (see Additional File 5).

The analyses were carried out using the STATA^® (Statutory, College Station, Texas) statistical package, except for the multi-level analysis that was carried out using R (RDC Team; http://​www.​R-project.​org).

First of all, our study was observational, retrospective and used administrative databases as the source of information. While criteria for inclusion tried to homogenize the sample and minimize data quality problems, the shortcuts and pitfalls of this data for risk-adjustment are well known [43]. Regarding the main diagnosis, a recent study in the Netherlands has shown that hip fracture may be miscoded by up to the 3% in administrative databases [44]. The paucity of secondary diagnoses recorded (two-third of patients with 3 or less) and the high volume of patients with a Charlson score equal to 0 (our sample only included elderly people, more than 50% over 80 years-old) warns about the possibility of information biases that might affect risk-adjustment. Although Spanish regional departments of health have carried out some internal audits on the Minimum Basic Data Sets in their respective regions, there are very few published studies on data quality in Spain (to our knowledge none of them auditing the hip fracture coding) and its results show problems of accuracy, completeness and information biases [45‐47] which are expected to have affected the risk adjustment analysis in some way. Nonetheless, and for the main objective of our study, these problems would suppose an important bias only if there was differential between early or delayed surgery groups and/or with the main study endpoint. As observed in the stratified analysis (Table 4), differences in mortality according to coding strata were not significant, except in the case of patients with more than 6 secondary diagnoses where the death rate was higher in early surgery (a 2.54% absolute difference).

Also related to information biases is the fact that the MBDS do not provide chronological information about secondary diagnoses. While diagnoses for chronic conditions can be confidently treated as comorbidities, acute conditions could be comorbidities or complications. The inability to distinguish among them, together with the possibility of a selective recording of a severe diagnosis depending on the evolution of the patient, reduces the capacity to properly adjust for individual risk. The effect on the estimated differences would be expected to be small, since no differences in mortality were found between any one of the strata, either in the Charlson or RMI Indexes (Table 4). However, the fact that some recorded complications could be either the cause or consequence of delayed surgery does not allow the ruling-out of biased estimates on mortality in the case of this surgical approach. These information biases could tend to overadjust cases with poor outcomes and to attenuate (if they really exist) mortality differences between early and delayed surgery.

The use of inpatient death as the main endpoint (vs. mortality within a fixed time, such as 30 days or one-year from admission, data not available in our study) involves several inconveniences [48]. More aggressive discharge policies can reduce the "mortality" rates if patients are discharged "alive" to die at home. On the contrary, a longer length of stay (LOS) increases the probability of identifying some outcomes such as patient mortality. Because delayed surgery is associated with longer LOS the mortality in this group is probably overemphasized. If poor patient condition was associated with the decision to delay the intervention then these problems could also contribute to overstating mortality in this group. Additionally, the recording of a secondary diagnosis could be associated with LOS and poor outcomes, also affecting the covariates adjustment. The available endpoint does not permit the overcoming of these problems that could bias our results if different discharge policies are used in patients with poor outcomes (associated with one or another of the surgical alternatives evaluated).

One limitation that could potentially affect the estimation of non-differences between early and delayed surgery is the lack of information on the probability of a patient dying in the days immediately following admission. Moreover, the observed cutback in the total length of stay (one day, in spite of the increase in pre-operative stays) over the follow-up period may increase the risk of bias. If we could assume that mortality after discharge followed a random distribution, the estimates would not be affected. We have favoured this possibility by adding "in extremis" patients (i.e. patients who are discharged to die at home) to the case definition, but we are not able to rule out that the hospitals, in our sample, discharge patients depending on the risk of short-term death, thus affecting the non-difference on the timing of surgery.

Finally, the inclusion of only 8 autonomous communities could also be a source of bias if the non-included communities showed different behaviour. Nevertheless, the participating autonomous communities are quite diverse and include practically half of the Spanish population.

Late hip fracture surgery could be due to different reasons. Unavoidable delays in more severe patients unfit for surgery, better control of hemorrhagic risks in patients taking oral anticoagulants or antiplatelet agents, and organizational problems are the most important, being problems caused by the shortage of orthopaedic surgeons or anaesthesiologists very unusual in Spain. With some uncertainty derived from the limitations stated, our results do not show any independent association between early surgery and reduced mortality and does not support the use of early surgery as a quality indicator related with poor outcomes in the Spanish NHS setting. Anyway, and independently of its impact on mortality, while early surgery could shorten the length of stay and its more comfortable for patients, we found a huge variation in the proportion of "early surgery" between hospitals and a surprisingly low average of "early surgery" compared to international standards. These data suggest the need for quality improvement in Spain in terms of good practices in identifying patients that can (or cannot) be operated early. On the other hand, and beyond the delays for organizational reasons, orthopaedic surgeons and/or anaesthesiologists seem to use specific criteria for delaying certain patients. Probably, before developing new studies that analyse the effectiveness of early vs. delayed surgery in all hip fracture patients regardless their clinical status, new research projects should be directed to identify patients with more probability of benefitting from the timing of each surgical approach in the actual context of care for hip fracture patients.