Trauma registry record linkage: methodological approach to benefit from complementary data using the example of the German Pelvic Injury Register and the TraumaRegister DGU®

The Pelvic Injury Register (PIR)

The TraumaRegister DGU® (TR)

The TR was founded in 1993 as a prospective, multicentre, standardized and anonymous documentation of multiple injured trauma patients. Data are collected at four consecutive post-trauma phases from injury to hospital discharge: (i.) pre-hospital phase; (ii.) emergency room/department and initial surgery (until admission to the intensive care unit (ICU)); (iii.) ICU and (iv.) outcome status at discharge and description of injuries and procedures takes place. The registry contains detailed information on demographics, injury pattern, comorbidities, pre- and in-hospital management, time course, relevant laboratory findings including data on transfusion, and outcome of each individual (Table 1). Fluid management is addressed, including records of the infusion volume of crystalloids and colloids started preclinically as well as the blood products transfused during the first 24 hours in hospital. The predominant inclusion criterion is that the injury severity or trauma load requires the patient to stay in the ICU. All injuries are coded using the Abbreviated Injury Scale (AIS). The majority of contributing hospitals are level I trauma centers. In 2009, there were 242 hospitals affiliated with the TR, with an upward trend. The TR is a voluntary register and it is approved by the review board of the German Trauma Society and is in compliance with the institutional requirements. In accordance to the PIR, the TR is a voluntary and anonymous registry that needs no patient consent.

The linkage procedure of the two data bases

To identify identical patients entered in both registries, a data set was prepared containing the data of all hospitals contributing to both registries during the period 2004-2009. The data set of the TR was then reduced to cases with injuries coded within a specific AIS (1998) numerical injury identifier, i.e. 856xxx.x, containing pelvic ring and acetabular fractures [12]. In the next step, the two data sets were limited to cases where specific hospitals contributed to both registries during the same year. The data sets from the registries were then imported via .csv files into SPSS software (SPSS, Chicago, IL, USA). The unique hospital code used in the TR was introduced into the PIR data set, after which a match-code was created containing the hospital code, dates of admission and discharge, and the age and sex of the patient. Before linking the data sets, each set was checked for possible duplicates. After controlling and (if required) deleting duplicates, the match-code was used to link the two data sets and merged the hits in a new file. This new data file contained the variables of both registries for each patient.

Checking for data concordance

The following variables were chosen to investigate the concordance of data entries in the two registries: initial haemoglobin and blood pressure levels on arrival at the emergency department, the ISS, and mortality. Records were considered a match if they were identical (the gold standard). Because of the slightly different methods for coding the ISS within the two registries, however, a discrepancy of ± 9 points was tolerated. Also, we accepted discrepancies of ± 1 g/dl for the haemoglobin level and ± 10 mmHg for blood pressure. We considered that there was no substantial impact on clinical treatment or outcome within these ranges.

Statistical analysis

Data are presented as the mean and standard deviation for continuous variables and as a percentage for categorical variables. The percentage of exact matches and percentage of matches within the tolerated range are given for each variable. Differences in documentation were compared using Wilcoxon test for continuous variables and Chi² for categorical variables. Statistics were calculated using SPSS Statistical Software Package Version 19 (SPSS Inc., Chicago, USA). A p value of <0.05 was considered statistically significant.

Linkage rate

In multiple trauma patients, the incidence of concomitant pelvic fractures increases up to more than 25% [1–3]. Vice versa, up to 89% of patients sustaining high-energy derived pelvic fractures have at least one associated injury [13, 14]. We developed an approach to find patients whose data had been entered into two anonymous trauma registries. Using this approach we identified 420 multiple trauma patients with concomitant pelvic fractures documented in both the PIR and the TR. Prima facie, the linkage rates for the intersection set of 15.7% in the PIR and 44.4% in the TR appear small, which means that only about one in every six patients with a pelvic fracture enrolled in the PIR fulfilled the TR inclusion criterion of “care in the ICU”. On the other hand, approximately every second patient with a pelvic fracture enrolled in the TR was also documented in the PIR. The reason for this discrepancy is that enrolment is mainly determined by the particular emphasis of each registry and the study populations themselves.

Starting with the PIR, the drop-outs might be explained by the large number of isolated acetabular and benign anterior pelvic ring fractures (type A according to the Tile/OTA classification) caused by low-energy trauma in the very elderly population [7, 10]. Each of these injuries is documented in the PIR but not routinely in the TR because only in rare cases are these patients treated initially as emergency cases in the ICU. Comparing the mean age and the fracture distribution of solely PIR data from Pohlemann et al. with our linked data set, the higher average in age (54.4 years vs. 41.6 years) as well as the higher percentage of type A fractures (42.8% vs. 19.8%) in consequence is found in the PIR [10]. In contrast, the reasons for the even lower drop-outs in the TR are more speculative. Finally, the match code itself could explain the large number of dropouts as there was no tolerance given (e.g. in admission and discharge dates), reflecting an ad hoc deterministic linkage. In contrast to the recently published use of linkage procedures based on encrypted personal identifiers or use of a probabilistic record linkage [15, 16], an extensive hand search of both anonymous data sets might have offered the possibility of isolating patients’ data that were mistakenly omitted from one registry and/or to answer the question of how many persons were missed by our linkage criteria. Because of the huge efforts in time and resources, we waived a thorough hand search. This matter, in fact, represents an explicit limitation of our study.

Published articles have compared administrative databases for hospital reimbursement. Trauma registry databases are available for research on epidemiology, practice variation, and outcome [17, 18]. There have also been inter-registry analyses in the field of trauma [19, 20]. Although the literature is filled with comparisons of these various medical registries, this is the first time that linking data from two clinical trauma registries to create an enlarged data set for each patient has been described. It is also the first time that the benefit of these complementary data has been addressed.

Data validation

Looking for data validation, we have to address the differences between data completeness and data concordance. Because the ISS and mortality records are directly or indirectly mandatory in both databases, not surprisingly the completion rate was 100% for both parameters in both registries. In contrast, there was a poor completion rate (17%, 73/420) for the initial blood pressure and haemoglobin measurements in the PIR. Usually, this rate is unacceptable compared with the 99% reported in the literature [21]. It should be kept in mind, however, that both parameters are mandatory only for complex pelvic injuries. This appears to be emphasized for the PIR only, reflecting the insistence for vital parameters or the initial haemoglobin level only for complex pelvic injuries. Nevertheless, a high completion rate of the collecting data should be in the interest of each registry, and future data validation schemes must be implemented in the PIR.

The results for data concordance also appear inconsistent. That is, statistics showed no significant differences for the ISS or mortality, but this was not true for the initial blood pressure and haemoglobin measurements. Undoubtedly, the haemoglobin level and hemodynamics of trauma patients influence their outcome and survival. Using only PIR-derived data, Holstein et al. investigated predictors of mortality in patients with pelvic fractures. They found out that the median haemoglobin level of the “survivor group” was 10 g/dl compared with 7 g/dl in the “nonsurvivor group” [8]. Regarding the initial haemoglobin levels of our linkage groups, both mean haemoglobin levels were found to be just between the above-mentioned extremes (8.6 ± 2.9 g/dl from PIR data and 9.6 ± 3.1 g/dl from TR data). Statistics nevertheless showed significant differences, reflecting inconsistency for the data entry level of both clinical parameters. Potential explanations for this inconsistency are numerous (e.g. data entry or coding errors; rounding up or down of numerical values; replicated values during the same treatment period; different error deviations of different haematology analysers or blood pressure-measuring devices).

Finally, by using the defined tolerance levels of these two parameters in the PIR and the TR, we reached the same data entry error rates as in the National Trauma Data Bank (19-76%), although for different records [22–24]. Data quality is still an important challenge for the PIR and the TR. Our results confirmed the need for establishing appropriate validation protocols in the future [24, 25].

Future perspectives and confidentiality concerns

We will also address the use of a unique patient identifier in registers like social security number, which is allowed in some European countries, but not allowed in Germany due to ethical considerations and confidentiality concerns. Such an identifier may considerably help identify a patient across registries, thereby allowing evaluation of the patient’s course through the entire health system. An example is the MEMDoc® module system used in the French hip registry and in other registries that use the social security number even over country borders [26].