Get 20M+ Full-Text Papers For Less Than $1.50/day. Start a 14-Day Trial for You or Your Team.

Learn More →

Record linkage is feasible with non‐identifiable trauma and rehabilitation datasets

Record linkage is feasible with non‐identifiable trauma and rehabilitation datasets disability in adults of working age, R Objectives: 1) Describe probabilistic linkage (PL) for road trauma and rehabilitation records in particularly in people aged 15-29 Capturing the rehabilitation New South Wales (NSW) Australia. 2) Determine the accuracy of linkage for these records. years of age. needs and outcomes of road trauma Methods: Data were extracted from the NSW Trauma Registry for all road trauma admissions patients is critical for service planning and for the years 2009–2012 and from Australasian Rehabilitation Outcomes Centre for January for determining the cost effectiveness and 2009 to June 2013. PL was performed using: age; sex; residential postcode; and date of acute quality of service provision. discharge = date of admission to rehabilitation. False matches were cases that linked but were The NSW Trauma Registry contains de- not true matches; they were determined by manual review. Reasons for incomplete linkages identified trauma data collected from were explored. The benefits and limitations of the linked study dataset are described. metropolitan and regional trauma centres in Results: Of 3,256 road trauma records, 683 were matched to rehabilitation records. Using the New South Wales. It is managed by the NSW field of ‘discharge destination’ from the trauma records, 265 patients with unmatched records Institute of Trauma and Injury Management were discharged to inpatient rehabilitation (missed matches). This gave an overall 72% linkage (NSW ITIM) and captures data only from the rate (or sensitivity) using PL. There were 16 cases of false matches, giving a specificity of 99%. acute care episodes with moderate to severe Conclusion: It was feasible to use PL to link road trauma and rehabilitation datasets in the injury. It is estimated that 28% of trauma absence of identifiers. However, this needed to be combined with careful manual review before patients required inpatient rehabilitation. the linked dataset could be used to make inferences on trauma rehabilitation outcomes. Data relating to this phase of care is not Implication: PL may be a cost-effective way to capture inpatient rehabilitation outcomes of captured by the NSW Trauma Registry. multi-trauma patients. It has been suggested that for trauma Key words: record linkage, probabilistic linkage, trauma, rehabilitation registries to compare improvements over time, it is important that survivors with poorer outcomes be identified and that the present Physicians. It is withing the Australian Health the linkage manually for 375 patients using measurement of discharge outcomes is Services Research Institute at the University of impairment, injury date and date of birth. inadequate for this purpose. There are no Wollongong in New South Wales. Manual linkage is only feasible for small published inpatient rehabilitation outcomes The AROC database contains de-identified datasets. For large datasets, deterministic for this group in Australia. Conducting data from the inpatient specialist medical or probabilistic linkage (PL) techniques are specific studies of the inpatient rehabilitation rehabilitation sector. The AROC database more suitable. Deterministic linkage involves outcomes following acute care is expensive contains nearly all rehabilitation episodes of linking records based on exact agreement of and resource intensive. care in Australia and New Zealand. the selected matching variables and would be Linkage of routinely collected administrative preferred if the two datasets had identifiers Data linkage between trauma and and clinical health data may provide an such as name, address or unique universal rehabilitation datasets from the US was first efficient alternative. 6 health insurance (Medicare) number. PL is attempted in 1996 by Copes et al. They The Australasian Rehabilitation Outcomes based on the probabilities of agreement and concluded that “linking records to create the Centre (AROC) was established in 2002 by disagreement between a range of matching study data base was arduous and could not the Australasian Faculty of Rehabilitation variables. This allows linkage of datasets be practically accomplished on a large scale Medicine of the Royal Australasian College of where one or both do not have identifiers. or on a continuing basis”. However, they did 1. St. Vincent’s Hospital, New South Wales 2. Liverpool Hospital, New South Wales 3. School of Public Health and Community Medicine, University of New South Wales 4. Australian Health Services Research Institute, University of Wollongong, New South Wales Correspondence to: Dr Jane Wu, St. Vincent’s Hospital, Sacred Heart Rehabilitation Service, 170 Darlinghurst Rd, Darlinghurst, NSW 2010; e-mail: Jane.Wu@svha.org.au Submitted: May 2015; Revision requested: August 2015; Accepted: November 2015 The authors have stated they have no conflict of interest. Aust NZ J Public Health. 2016; 40:245-9; doi: 10.1111/1753-6405.12510 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 245 © 2015 Public Health Association of Australia Wu et al. Due to recent advances in data linkage 2) residential postcode; 3) age; and 4) date To avoid duplicate records in the trauma technologies, Australian researchers have of discharge from acute care (episode begin dataset (such as those who present to one 7-8 published many projects using linked data. date of inpatient rehabilitation). trauma centre and are then transferred to Data linkage projects are more feasible another trauma centre), patients with a length The Centre for Health Record Linkages than ever before, with an increased capacity of stay (LOS) of three days or fewer with (CHeReL) was an agency that provided to effectively combine information across unlinked records were manually reviewed, researchers technical assistance with data distinct and large-sized data sources. This and those who were recorded as having been linkage. Custodians of the NSW Trauma was, however, the first attempt to link NSW transferred to another acute care facility or Registry and AROC provided CHeReL with an Trauma Registry data with AROC data. with unknown disposition were excluded encrypted source record number and the four from analysis. This process was thought to The objectives of this study were to: describe matching variables for each individual in their be necessary to avoid skewing outcomes PL for road trauma and rehabilitation records dataset. CHeReL linked these records using (e.g. a critically injured patient who spends in NSW, and determine the accuracy of probabilistic matching of the four variables, one day at a regional trauma centre before linkage for these records. assigned a project-specific person number being moved to a metropolitan trauma centre (PPN) for each person in the linked dataset will have a record with only one day as the and provided this to the data custodians. Methods total LOS in the first admission and a longer admission in the second trauma centre). Data provision to the researchers Data extraction The study was reviewed and approved by Criteria for inclusion in the NSW Trauma Data concatenation NSW Population & Health Services Research Registry included: 1) admission to a trauma Conceptually, the total contiguous period Ethics Committee (HREC/13/CIPHS/55). service in NSW within 14 days of injury; 2) spent in one or more hospitals after initial injuries classified as moderate, serious or The data custodian decrypted the source entry to hospital represented a single critical (i.e. an Injury Severity Score or ISS of 13 record number and merged the PPN with the patient journey after trauma. It was not or more); 3) death in hospital (irrespective of clinical variables that were approved for use uncommon for patients to be transferred to the ISS); and/or 4) admission to intensive care in the study. The source record number was an acute hospital during rehabilitation, and unit (irrespective of the ISS). removed and the researchers were provided then be subsequently transferred back to The NSW Trauma Registry collected with the PPN and the clinical information. rehabilitation. Individual episodes of care mechanism of injury in their dataset. Road The researchers were then able to combine were combined to accurately reconstruct trauma was defined as any injury occurring the records for the same person from the the complete journey of rehabilitation care. on NSW roads involving a moving vehicle. It different datasets using the PPN. AROC This process, called data concatenation, was included drivers and passengers in cars, buses also provided the researchers repeated necessary to avoid erroneous LOS and final and heavy vehicles, cyclists, motorcyclists, records (different rehabilitation episodes of outcome and errors in the calculation of cost pedestrians and others (such as motorised care) of the same patient to allow for data of injury. scooters). Data were extracted from the NSW concatenation (see below). Acute care interruption (or gap between Trauma Registry for the calendar years of rehabilitation episodes) was defined as 2009–2012 (inclusive) using these variables: Manual review the interval from hospital separation in 1) road trauma as the mechanism of injury; To improve data quality and integrity, manual rehabilitation episode 1 until hospital entry and 2) age 18 and above. Patients who died in review of the two datasets was required. This in rehabilitation episode 2. The researchers hospital were excluded. was also done to standardise data, exclude defined this gap as <15 days. For those with Data were extracted from the AROC database unwanted records and check for accuracy of 15 days or more separation, the two episodes for the period January 2009 to June 2013 for linkage. were only considered contiguous if episode all NSW private and public hospitals. Excluded As the purpose of doing this record linkage one did not have an end accommodation from this initial dataset were patients with was to review rehabilitation outcomes of listed as home, residential care or transitional impairment codes that were not relevant to those with ISS >12 (moderate, serious and care. trauma, such as arthritis, cardiac, congenital critical injuries), the records for those with Concatenated rehabilitation episodes had deformities and developmental deformities. ISS <13 were excluded from the final study revised LOS and Functional Independence Data from AROC included information dataset. Measure (FIM) changes (start details were about the rehabilitation service, the person The datasets provided to the researchers taken from the identified primary episode receiving the rehabilitation, details about contained records as a link, possible link or and end details were taken from the the rehabilitation episode of care and clinical non-link. Manual review was required for: identified final episode). characteristics. Outcome measures included 1) All possible links (trauma records linked to functional status measured by the Functional more than one rehabilitation records). Audit Independence Measure (FIM) score and 2) Possible false links (linked trauma records A small sample of data were independently discharge destination. where the discharge destination was not checked by another researcher for recorded as ‘inpatient rehabilitation’. consistency in clinical reasoning. Data linkage 3) All duplicate records (that were deleted The variables common to both databases from final dataset). that were used for matching included: 1) sex; 246 Australian and New Zealand Journal of Public Health 2016 vol . 40 no . 3 © 2015 Public Health Association of Australia Epidemiology Linkage of trauma and rehabilitation datasets Table 1. This data linkage attempt missed of where they were managed (major vs Results 187 general rehabilitation records, 60 brain regional trauma services), age, injury severity The researchers received 3,734 trauma injury rehabilitation records and 18 spinal (as measured by the ISS), need for admission records and 846 rehabilitation records. rehabilitation records. We used this data to to the intensive care unit, LOS in acute care Manual review and data cleaning was calculate the sensitivity (72%), specificity and need for specialised (brain and spinal performed by one researcher (JW) as (99%), positive predictive value (98%) and injury) rehabilitation (Table 3). Using these described in Figure 1. negative predictive value (90%) of our data measures of severity, it would be reasonable linkage attempt (Table 2). to presume that study dataset was unlikely to To allow the researchers to map the complete be a biased sample and that the missed links inpatient journey post-trauma as one linked The “discharge destination” recorded in the were missed randomly. record, the following data concatenation NSW Trauma Registry was accurate 72% of the process was undertaken. The patients time. Of the 667 linked cases, the discharge with multiple episodes of rehabilitation destination was recorded incorrectly in 189 Discussion (n=53) were identified by AROC and all trauma records as ‘other’, ‘acute care facility’, Feasibility of data linkage their episodes (n=121) were provided to ‘intermediate facility’ or even ‘home’. the researchers. Data concatenation was The NSW Trauma registry and AROC used The missed matches were compared with the performed on all 53 patients of the 667 an opt-out method of consent that allowed linked cases and there were no significant patients (8%) who went to rehabilitation. almost all patients with trauma and all differences between the two groups in terms To calculate an estimate of error if the dataset did not use concatenated episodes, we used Table 1: Discharge destination of the trauma records with no links. the difference in the number of days in LOS Discharge destination – trauma records with no links (n=2590) Frequency Percentage between the two methods. We calculated Inpatient rehabilitation 187 7.2% the total number of rehabilitation bed days Brain injury unit 60 2.3% without data concatenation to be 33,465. Spinal injury unit 18 0.7% With concatenation, this number increased to Acute care facility 132 5.1% 37,480, which represented a 12% increase. Home (community) 1846 71.3% To detect false matches, all linked trauma Residential care, jail or psychiatric facility 19 0.7% records that had a discharge destination Unrecorded or other 327 12.6% recorded as home, acute care, intermediate care facility or unknown were reviewed Figure 1: Algorithm for data linkage of trauma and rehabilitation datasets. (n=205). Contradictions in clinical data were used as rationale for determining the link to be false (n=16). All of these false links had a rehabilitation admission date that predated the date of injury or date of discharge from acute care separated by at least one month. Collaborative clinical evidence was also used; for example, a trauma record with no brain injury in the injuries descriptor being linked to a rehabilitation episode in a brain injury unit was considered a false match. A second researcher (SF) independently reviewed a random sample of the false matches (n=5) and a random sample of the data concatenation process (n=10) to assess for accuracy. This researcher was in agreement with the clinical decisions or assumptions used for this manual review process. After this process of manual review and data cleaning, the final study dataset contained 3,273 road trauma records with 683 rehabilitation records; 667 of the rehabilitation records were true matches and 16 false matches. Of the trauma records with no links (n=2,590), the discharge destinations as recorded in the NSW Trauma Registry are displayed in 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 247 © 2015 Public Health Association of Australia Wu et al. rehabilitation episodes of care to be captured, However, our two datasets only had four rehabilitation but where those hospitals did therefore minimising bias. Opt-out rates are definite common variables that were used for not submit to AROC may account for some not recorded by the two data custodians this study. The only other variable that may missed links. Some of these inpatient services (personal communication) but the Victorian have been used from the trauma dataset was would have included district hospitals or State Trauma Registry, which has used a “date of injury”, which may be matched with small private hospital that provided ‘slow similar opt-out method, has reported that “date of onset of the impairment” from the stream’ rehabilitation where there was some <0.1% of the cases requested to opt out. rehabilitation dataset. allied health input, but not to the intensity that allowed it to be defined as rehabilitation. Lack of identifiable data made data linkage The ‘gold standard’ for identifying “true- Trauma nurses who collected and coded the difficult but not impossible. This study has positives” was to do a manual review of data, however, were probably unaware of this shown that out of the 932 patients recorded inpatient charts. However, there were no difference in terminology. to have been transferred to inpatient identifiers from the NSW Trauma Registry rehabilitation, 667 (72%) were able to be to allow an audit of true positives. Therefore Some missed links could be related to successfully identified using data linkage. Of the discharge destination variable was patients who were transferred interstate for the rehabilitation records that were missing, the best available gold standard against their rehabilitation. It was the researchers’ the study demonstrated that these records which matched or unmatched records were decision to limit the initial data linkage to were missed randomly. considered ‘true’ or ‘false’ . NSW data only, so that the size of project was smaller. If we included rehabilitation episodes The accuracy of the coding of the ‘discharge Accuracy of data linkage outside of NSW, the cost of this exercise destination’ in the trauma records was an was likely to be higher (due to increased The linkage rate of 72% could be considered issue in itself. This study found that this field complexity of linking a much larger number acceptable, but there were no reference was inaccurately coded 28% of the time. of files) and could potentially have led to an trauma data linkage that could be used for This finding may provide an impetus for increased number of false matches. comparison. the custodian of the NSW Trauma Registry to consider doing audits with trauma Using age as a matching variable was also In a recent systematic review, da Silveira 10 coordinators who are currently responsible problematic if a patient had a birthday during and Artmann (2009) concluded that the for data collection, coding and entry. the acute admission. Given that the average accuracy of probabilistic record linkage acute LOS was more than one month, at of databases ranged from 74% to 98% Reasons for missed matches least one in 12 would be one year older in sensitivity. This was based on 33 studies, rehabilitation records compared to the age but the authors did not specify whether We speculate that there were a number of recorded in the trauma records. If the NSW these studies had unique identifiers to use reasons for the loss of linkages. Some patients Trauma Registry (like AROC) collected dates for matching. The differences in linkage were transferred to their local acute hospital of birth rather than age, linkage rates would rates in various studies were dependent on or to residential aged care facility for respite have been higher. both the data quality, the specificity of the before going to rehabilitation, and therefore Quality of data in health information systems linkage variables and the number of available discharge date from acute care would not affects linkage rate. To date, there have not been matching variables that can be used. have matched the rehabilitation admission any audits to define data quality from either date. Increasing the number of variables for datasets. Poor data quality could have resulted matching could have increased the sensitivity. Patients who were transferred to hospitals for from clerical errors, missing information (in variables or even non-submission of an episode of care) and errors in coding. Clerical errors Table 2: Outcomes from data linkage. could have occurred when entering data, e.g. Records truly are from the same person Records truly are not from the same person a single digit error in the postcode would have Records matched Truly matched (TM) = 667 Falsely matched (FM) = 16 resulted in a missed link. Records not matched Falsely unmatched (FU) or missed matches = Truly unmatched (TU) = Although the participating rehabilitation 187 + 60 + 18 = 265 132 + 1,846 + 19 = 1,997 units were supposed to submit all patient The above numbers were used to calculate (1) sensitivity = TM/(TM+FU) = 72%, (2) specificity = TU/(TU+FM) = 99%, (3) positive predictive value = TM/ records, AROC did not have the resources (TM+FM) = 98%, (4) negative predictive value = TU/(TU+FU) = 88%. to check the rate of compliance with record submission (personal information from Table 3: Similarities between true and missed matches. a b AROC). All episodes submitted to AROC were True matches (n=667) Missed matches (n=265) p routinely audited at the point of submission Admitted to major trauma services 631 (95%) 254 (96%) 0.43 (as opposed to regional trauma services) and an audit report of the submitted episodes Age (years, SD) 50 (±22) 48 (±21) 0.15 would have been provided to the facility. Injury Severity Score (SD) 27 (±11) 27 (±11) 0.58 However, the onus was on the facility to make Required intensive care admission 465 (70%) 177 (68%) 0.39 corrections and resubmit where errors were Acute length of stay (days, SD) 40 (±26) 37 (±41) 0.26 found and/or data were missing. Rates of Transferred to specialised rehabilitation 233 (35%) 78 (29%) 0.11 resubmission were not available. (brain injury and spinal injury) a: Trauma records with linked rehabilitation records b: Trauma records with a discharge destination of inpatient rehabilitation but were not linked to rehabilitation records 248 Australian and New Zealand Journal of Public Health 2016 vol . 40 no . 3 © 2015 Public Health Association of Australia Epidemiology Linkage of trauma and rehabilitation datasets This study detected errors in coding in Benefits of this study dataset Acknowledgements discharge destination that affected data The final linked dataset will allow researchers The researchers are grateful to the Motor quality. This could provide an impetus for the to look at the inpatient rehabilitation Accidents Authority of NSW for funding this NSW Trauma Registry to review its coding outcomes of road trauma. It will allow project. We thank CheReL for doing the data system. Both registries could consider having for attempts to develop some models for linkage and the data custodians for the NSW data quality assurance processes in place and estimating functional gains, length of stay in Institute of Trauma and Injury Management could also publish their frequency of checks. rehabilitation, rehabilitation costs based on and AROC for extracting and recoding the An estimate of error from these audits is impairment group and FIM on rehabilitation data for the study dataset. important as a proxy measure of data quality admission. and therefore its usefulness. Although the sensitivity was only 72%, the At all levels, maintenance of a reliable registry References final dataset has 667 valid matches, which such as NSW Trauma Registry is very costly. will make it the largest cohort in the trauma 1. NSW Institute of Trauma and Injury Management. The Linking the NSW Trauma Registry with the 4 NSW Registry Profile of Serious to Critical Injuries, 2008. rehabilitation literature. This study has Sydney (AUST): New South Wales Health; 2011 [cited AROC dataset at various time periods may be demonstrated that the missed links were 2016 Mar 4]. Available from: http://www.aci.health. one option for providing valuable functional nsw.gov.au/__data/assets/pdf_file/0007/195334/ missed randomly. As both datasets were The_NSW_Trauma_Registry_Profile_of_Serious_to_ information for the 29% of multi-trauma population based, the data would provide Critical_Injuries_2008.pdf patients that require inpatient rehabilitation. 2. Dinh MM, Bein KJ, Gabbe BJ, Byrne CM, Petchell J, Lo valid answers to research questions on S, et al. A trauma quality improvement programme As trauma systems in NSW mature and trauma rehabilitation outcomes in NSW. associated with improved patient outcomes: 21 years survival rates increase, this information of experience at an Australian Major Trauma Centre. The process of data concatenation was would be reassuring if one can demonstrate Injury. 2014;45:830-4. considered an important benefit of this 3. Gabbe BJ, Simpson PM, Sutherland AM, Williamson that rehabilitation outcomes are improving OD, Judson R, Kossmann T, et al. Functional measures study dataset. AROC used the following and/or the proportion of patients who are at discharge. Are they useful predictors of longer business rules for concatenation: number of term outcomes for trauma registries? Ann Surg. discharged from rehabilitation with severe days between episodes being 0–14 days for 2008;247:854-9. disability is static or reducing. 4. Khan F, Amatya B, Hoffman K. Systematic review of spinal and 0–7 days for all other impairments multidisciplinary rehabilitation in patients with multiple (personal communication). We felt that these trauma. Br J Surg. 2012;99 Suppl 1:88-96. 5. Simmonds FD, Stevermuer TL. The AROC Annual rules were overly restrictive and, if applied to Conclusion Report: The State of Rehabilitation in Australia in this dataset, would only concatenate 32 sets 2012 [Internet]. Wollongong (AUST ): Australian This study has demonstrated that it is feasible Health Services Research Institute Australasian of patient data rather than 53 sets based on to link road trauma and rehabilitation Rehabilitation Outcomes Centre; 2013 [cited 2016 Mar our clinical reasoning. 4]. Available from: http://ahsri.uow.edu.au/content/ registries using non-identifiable data. groups/public/@web/@chsd/@aroc/documents/doc/ Our paper provided an alternative method However, this then needs to be combined uow160486.pdf for data concatenation for future database 6. Copes WS, Stark MM, Lawnick MM, et al. Linking data with a careful manual review of the results to from national trauma and rehabilitation registries. J studies involving AROC. This study has also make inferences on individual cases. Overall, Trauma. 1996;40(3):428-36. found that, if using concatenated data, the 7. Holman CDJ, Bass AJ, Rosman DL, et al. A decade of this process generated 72% valid matches. real cost of inpatient trauma rehabilitation data linkage in Western Australia: Strategic design, The missed links were missed randomly. applications and benefits of the WA data linkage (based on LOS) would be at least 12% more. This study shows that probabilistic linkage system. Aust Health Rev. 2008;32(4):766-77. 8. Glasson EJ, Hussain R. Linked data: Opportunities and with trauma and rehabilitation datasets is an challenges in disability research. J Intellect Dev Disabil. Future directions option when mapping of the flow of patients 2008;33(4):285-91. 9. Cameron PA, Finch CF, Gabbe BJ, et al. Developing Linkage studies are becoming more prevalent through multiple phases of care in a trauma Australia’s first statewide trauma registry: What are the due to improved methods and are considered system is required. lessons? Aust N Z J Surg. 2004;74(6):424-8. important in the field of injury and disability 10. Da Silveira DP, Artmann E. Accuracy of probabilistic record linkage applied to health databases: Systematic research. However, the quality of linked review. Rev Saude Publica. 2009;43(5):1-7. datasets will depend on the rate of linkage 11. Clarke DE. Practical introduction to record linkage for injury research. Inj Prev. 2004;10:186-91. and the quality of the data used. The rate of linkage could be improved if the NSW Trauma Registry collected unique identifiers. However, if privacy and/or ethical constraints prevent them collecting unique identifiers, they could consider collecting date of birth instead of age. 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 249 © 2015 Public Health Association of Australia http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Australian and New Zealand Journal of Public Health Wiley

Record linkage is feasible with non‐identifiable trauma and rehabilitation datasets

Loading next page...
 
/lp/wiley/record-linkage-is-feasible-with-non-identifiable-trauma-and-gE10uIyRgq

References (15)

Publisher
Wiley
Copyright
© 2016 Public Health Association of Australia
ISSN
1326-0200
eISSN
1753-6405
DOI
10.1111/1753-6405.12510
pmid
27028098
Publisher site
See Article on Publisher Site

Abstract

disability in adults of working age, R Objectives: 1) Describe probabilistic linkage (PL) for road trauma and rehabilitation records in particularly in people aged 15-29 Capturing the rehabilitation New South Wales (NSW) Australia. 2) Determine the accuracy of linkage for these records. years of age. needs and outcomes of road trauma Methods: Data were extracted from the NSW Trauma Registry for all road trauma admissions patients is critical for service planning and for the years 2009–2012 and from Australasian Rehabilitation Outcomes Centre for January for determining the cost effectiveness and 2009 to June 2013. PL was performed using: age; sex; residential postcode; and date of acute quality of service provision. discharge = date of admission to rehabilitation. False matches were cases that linked but were The NSW Trauma Registry contains de- not true matches; they were determined by manual review. Reasons for incomplete linkages identified trauma data collected from were explored. The benefits and limitations of the linked study dataset are described. metropolitan and regional trauma centres in Results: Of 3,256 road trauma records, 683 were matched to rehabilitation records. Using the New South Wales. It is managed by the NSW field of ‘discharge destination’ from the trauma records, 265 patients with unmatched records Institute of Trauma and Injury Management were discharged to inpatient rehabilitation (missed matches). This gave an overall 72% linkage (NSW ITIM) and captures data only from the rate (or sensitivity) using PL. There were 16 cases of false matches, giving a specificity of 99%. acute care episodes with moderate to severe Conclusion: It was feasible to use PL to link road trauma and rehabilitation datasets in the injury. It is estimated that 28% of trauma absence of identifiers. However, this needed to be combined with careful manual review before patients required inpatient rehabilitation. the linked dataset could be used to make inferences on trauma rehabilitation outcomes. Data relating to this phase of care is not Implication: PL may be a cost-effective way to capture inpatient rehabilitation outcomes of captured by the NSW Trauma Registry. multi-trauma patients. It has been suggested that for trauma Key words: record linkage, probabilistic linkage, trauma, rehabilitation registries to compare improvements over time, it is important that survivors with poorer outcomes be identified and that the present Physicians. It is withing the Australian Health the linkage manually for 375 patients using measurement of discharge outcomes is Services Research Institute at the University of impairment, injury date and date of birth. inadequate for this purpose. There are no Wollongong in New South Wales. Manual linkage is only feasible for small published inpatient rehabilitation outcomes The AROC database contains de-identified datasets. For large datasets, deterministic for this group in Australia. Conducting data from the inpatient specialist medical or probabilistic linkage (PL) techniques are specific studies of the inpatient rehabilitation rehabilitation sector. The AROC database more suitable. Deterministic linkage involves outcomes following acute care is expensive contains nearly all rehabilitation episodes of linking records based on exact agreement of and resource intensive. care in Australia and New Zealand. the selected matching variables and would be Linkage of routinely collected administrative preferred if the two datasets had identifiers Data linkage between trauma and and clinical health data may provide an such as name, address or unique universal rehabilitation datasets from the US was first efficient alternative. 6 health insurance (Medicare) number. PL is attempted in 1996 by Copes et al. They The Australasian Rehabilitation Outcomes based on the probabilities of agreement and concluded that “linking records to create the Centre (AROC) was established in 2002 by disagreement between a range of matching study data base was arduous and could not the Australasian Faculty of Rehabilitation variables. This allows linkage of datasets be practically accomplished on a large scale Medicine of the Royal Australasian College of where one or both do not have identifiers. or on a continuing basis”. However, they did 1. St. Vincent’s Hospital, New South Wales 2. Liverpool Hospital, New South Wales 3. School of Public Health and Community Medicine, University of New South Wales 4. Australian Health Services Research Institute, University of Wollongong, New South Wales Correspondence to: Dr Jane Wu, St. Vincent’s Hospital, Sacred Heart Rehabilitation Service, 170 Darlinghurst Rd, Darlinghurst, NSW 2010; e-mail: Jane.Wu@svha.org.au Submitted: May 2015; Revision requested: August 2015; Accepted: November 2015 The authors have stated they have no conflict of interest. Aust NZ J Public Health. 2016; 40:245-9; doi: 10.1111/1753-6405.12510 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 245 © 2015 Public Health Association of Australia Wu et al. Due to recent advances in data linkage 2) residential postcode; 3) age; and 4) date To avoid duplicate records in the trauma technologies, Australian researchers have of discharge from acute care (episode begin dataset (such as those who present to one 7-8 published many projects using linked data. date of inpatient rehabilitation). trauma centre and are then transferred to Data linkage projects are more feasible another trauma centre), patients with a length The Centre for Health Record Linkages than ever before, with an increased capacity of stay (LOS) of three days or fewer with (CHeReL) was an agency that provided to effectively combine information across unlinked records were manually reviewed, researchers technical assistance with data distinct and large-sized data sources. This and those who were recorded as having been linkage. Custodians of the NSW Trauma was, however, the first attempt to link NSW transferred to another acute care facility or Registry and AROC provided CHeReL with an Trauma Registry data with AROC data. with unknown disposition were excluded encrypted source record number and the four from analysis. This process was thought to The objectives of this study were to: describe matching variables for each individual in their be necessary to avoid skewing outcomes PL for road trauma and rehabilitation records dataset. CHeReL linked these records using (e.g. a critically injured patient who spends in NSW, and determine the accuracy of probabilistic matching of the four variables, one day at a regional trauma centre before linkage for these records. assigned a project-specific person number being moved to a metropolitan trauma centre (PPN) for each person in the linked dataset will have a record with only one day as the and provided this to the data custodians. Methods total LOS in the first admission and a longer admission in the second trauma centre). Data provision to the researchers Data extraction The study was reviewed and approved by Criteria for inclusion in the NSW Trauma Data concatenation NSW Population & Health Services Research Registry included: 1) admission to a trauma Conceptually, the total contiguous period Ethics Committee (HREC/13/CIPHS/55). service in NSW within 14 days of injury; 2) spent in one or more hospitals after initial injuries classified as moderate, serious or The data custodian decrypted the source entry to hospital represented a single critical (i.e. an Injury Severity Score or ISS of 13 record number and merged the PPN with the patient journey after trauma. It was not or more); 3) death in hospital (irrespective of clinical variables that were approved for use uncommon for patients to be transferred to the ISS); and/or 4) admission to intensive care in the study. The source record number was an acute hospital during rehabilitation, and unit (irrespective of the ISS). removed and the researchers were provided then be subsequently transferred back to The NSW Trauma Registry collected with the PPN and the clinical information. rehabilitation. Individual episodes of care mechanism of injury in their dataset. Road The researchers were then able to combine were combined to accurately reconstruct trauma was defined as any injury occurring the records for the same person from the the complete journey of rehabilitation care. on NSW roads involving a moving vehicle. It different datasets using the PPN. AROC This process, called data concatenation, was included drivers and passengers in cars, buses also provided the researchers repeated necessary to avoid erroneous LOS and final and heavy vehicles, cyclists, motorcyclists, records (different rehabilitation episodes of outcome and errors in the calculation of cost pedestrians and others (such as motorised care) of the same patient to allow for data of injury. scooters). Data were extracted from the NSW concatenation (see below). Acute care interruption (or gap between Trauma Registry for the calendar years of rehabilitation episodes) was defined as 2009–2012 (inclusive) using these variables: Manual review the interval from hospital separation in 1) road trauma as the mechanism of injury; To improve data quality and integrity, manual rehabilitation episode 1 until hospital entry and 2) age 18 and above. Patients who died in review of the two datasets was required. This in rehabilitation episode 2. The researchers hospital were excluded. was also done to standardise data, exclude defined this gap as <15 days. For those with Data were extracted from the AROC database unwanted records and check for accuracy of 15 days or more separation, the two episodes for the period January 2009 to June 2013 for linkage. were only considered contiguous if episode all NSW private and public hospitals. Excluded As the purpose of doing this record linkage one did not have an end accommodation from this initial dataset were patients with was to review rehabilitation outcomes of listed as home, residential care or transitional impairment codes that were not relevant to those with ISS >12 (moderate, serious and care. trauma, such as arthritis, cardiac, congenital critical injuries), the records for those with Concatenated rehabilitation episodes had deformities and developmental deformities. ISS <13 were excluded from the final study revised LOS and Functional Independence Data from AROC included information dataset. Measure (FIM) changes (start details were about the rehabilitation service, the person The datasets provided to the researchers taken from the identified primary episode receiving the rehabilitation, details about contained records as a link, possible link or and end details were taken from the the rehabilitation episode of care and clinical non-link. Manual review was required for: identified final episode). characteristics. Outcome measures included 1) All possible links (trauma records linked to functional status measured by the Functional more than one rehabilitation records). Audit Independence Measure (FIM) score and 2) Possible false links (linked trauma records A small sample of data were independently discharge destination. where the discharge destination was not checked by another researcher for recorded as ‘inpatient rehabilitation’. consistency in clinical reasoning. Data linkage 3) All duplicate records (that were deleted The variables common to both databases from final dataset). that were used for matching included: 1) sex; 246 Australian and New Zealand Journal of Public Health 2016 vol . 40 no . 3 © 2015 Public Health Association of Australia Epidemiology Linkage of trauma and rehabilitation datasets Table 1. This data linkage attempt missed of where they were managed (major vs Results 187 general rehabilitation records, 60 brain regional trauma services), age, injury severity The researchers received 3,734 trauma injury rehabilitation records and 18 spinal (as measured by the ISS), need for admission records and 846 rehabilitation records. rehabilitation records. We used this data to to the intensive care unit, LOS in acute care Manual review and data cleaning was calculate the sensitivity (72%), specificity and need for specialised (brain and spinal performed by one researcher (JW) as (99%), positive predictive value (98%) and injury) rehabilitation (Table 3). Using these described in Figure 1. negative predictive value (90%) of our data measures of severity, it would be reasonable linkage attempt (Table 2). to presume that study dataset was unlikely to To allow the researchers to map the complete be a biased sample and that the missed links inpatient journey post-trauma as one linked The “discharge destination” recorded in the were missed randomly. record, the following data concatenation NSW Trauma Registry was accurate 72% of the process was undertaken. The patients time. Of the 667 linked cases, the discharge with multiple episodes of rehabilitation destination was recorded incorrectly in 189 Discussion (n=53) were identified by AROC and all trauma records as ‘other’, ‘acute care facility’, Feasibility of data linkage their episodes (n=121) were provided to ‘intermediate facility’ or even ‘home’. the researchers. Data concatenation was The NSW Trauma registry and AROC used The missed matches were compared with the performed on all 53 patients of the 667 an opt-out method of consent that allowed linked cases and there were no significant patients (8%) who went to rehabilitation. almost all patients with trauma and all differences between the two groups in terms To calculate an estimate of error if the dataset did not use concatenated episodes, we used Table 1: Discharge destination of the trauma records with no links. the difference in the number of days in LOS Discharge destination – trauma records with no links (n=2590) Frequency Percentage between the two methods. We calculated Inpatient rehabilitation 187 7.2% the total number of rehabilitation bed days Brain injury unit 60 2.3% without data concatenation to be 33,465. Spinal injury unit 18 0.7% With concatenation, this number increased to Acute care facility 132 5.1% 37,480, which represented a 12% increase. Home (community) 1846 71.3% To detect false matches, all linked trauma Residential care, jail or psychiatric facility 19 0.7% records that had a discharge destination Unrecorded or other 327 12.6% recorded as home, acute care, intermediate care facility or unknown were reviewed Figure 1: Algorithm for data linkage of trauma and rehabilitation datasets. (n=205). Contradictions in clinical data were used as rationale for determining the link to be false (n=16). All of these false links had a rehabilitation admission date that predated the date of injury or date of discharge from acute care separated by at least one month. Collaborative clinical evidence was also used; for example, a trauma record with no brain injury in the injuries descriptor being linked to a rehabilitation episode in a brain injury unit was considered a false match. A second researcher (SF) independently reviewed a random sample of the false matches (n=5) and a random sample of the data concatenation process (n=10) to assess for accuracy. This researcher was in agreement with the clinical decisions or assumptions used for this manual review process. After this process of manual review and data cleaning, the final study dataset contained 3,273 road trauma records with 683 rehabilitation records; 667 of the rehabilitation records were true matches and 16 false matches. Of the trauma records with no links (n=2,590), the discharge destinations as recorded in the NSW Trauma Registry are displayed in 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 247 © 2015 Public Health Association of Australia Wu et al. rehabilitation episodes of care to be captured, However, our two datasets only had four rehabilitation but where those hospitals did therefore minimising bias. Opt-out rates are definite common variables that were used for not submit to AROC may account for some not recorded by the two data custodians this study. The only other variable that may missed links. Some of these inpatient services (personal communication) but the Victorian have been used from the trauma dataset was would have included district hospitals or State Trauma Registry, which has used a “date of injury”, which may be matched with small private hospital that provided ‘slow similar opt-out method, has reported that “date of onset of the impairment” from the stream’ rehabilitation where there was some <0.1% of the cases requested to opt out. rehabilitation dataset. allied health input, but not to the intensity that allowed it to be defined as rehabilitation. Lack of identifiable data made data linkage The ‘gold standard’ for identifying “true- Trauma nurses who collected and coded the difficult but not impossible. This study has positives” was to do a manual review of data, however, were probably unaware of this shown that out of the 932 patients recorded inpatient charts. However, there were no difference in terminology. to have been transferred to inpatient identifiers from the NSW Trauma Registry rehabilitation, 667 (72%) were able to be to allow an audit of true positives. Therefore Some missed links could be related to successfully identified using data linkage. Of the discharge destination variable was patients who were transferred interstate for the rehabilitation records that were missing, the best available gold standard against their rehabilitation. It was the researchers’ the study demonstrated that these records which matched or unmatched records were decision to limit the initial data linkage to were missed randomly. considered ‘true’ or ‘false’ . NSW data only, so that the size of project was smaller. If we included rehabilitation episodes The accuracy of the coding of the ‘discharge Accuracy of data linkage outside of NSW, the cost of this exercise destination’ in the trauma records was an was likely to be higher (due to increased The linkage rate of 72% could be considered issue in itself. This study found that this field complexity of linking a much larger number acceptable, but there were no reference was inaccurately coded 28% of the time. of files) and could potentially have led to an trauma data linkage that could be used for This finding may provide an impetus for increased number of false matches. comparison. the custodian of the NSW Trauma Registry to consider doing audits with trauma Using age as a matching variable was also In a recent systematic review, da Silveira 10 coordinators who are currently responsible problematic if a patient had a birthday during and Artmann (2009) concluded that the for data collection, coding and entry. the acute admission. Given that the average accuracy of probabilistic record linkage acute LOS was more than one month, at of databases ranged from 74% to 98% Reasons for missed matches least one in 12 would be one year older in sensitivity. This was based on 33 studies, rehabilitation records compared to the age but the authors did not specify whether We speculate that there were a number of recorded in the trauma records. If the NSW these studies had unique identifiers to use reasons for the loss of linkages. Some patients Trauma Registry (like AROC) collected dates for matching. The differences in linkage were transferred to their local acute hospital of birth rather than age, linkage rates would rates in various studies were dependent on or to residential aged care facility for respite have been higher. both the data quality, the specificity of the before going to rehabilitation, and therefore Quality of data in health information systems linkage variables and the number of available discharge date from acute care would not affects linkage rate. To date, there have not been matching variables that can be used. have matched the rehabilitation admission any audits to define data quality from either date. Increasing the number of variables for datasets. Poor data quality could have resulted matching could have increased the sensitivity. Patients who were transferred to hospitals for from clerical errors, missing information (in variables or even non-submission of an episode of care) and errors in coding. Clerical errors Table 2: Outcomes from data linkage. could have occurred when entering data, e.g. Records truly are from the same person Records truly are not from the same person a single digit error in the postcode would have Records matched Truly matched (TM) = 667 Falsely matched (FM) = 16 resulted in a missed link. Records not matched Falsely unmatched (FU) or missed matches = Truly unmatched (TU) = Although the participating rehabilitation 187 + 60 + 18 = 265 132 + 1,846 + 19 = 1,997 units were supposed to submit all patient The above numbers were used to calculate (1) sensitivity = TM/(TM+FU) = 72%, (2) specificity = TU/(TU+FM) = 99%, (3) positive predictive value = TM/ records, AROC did not have the resources (TM+FM) = 98%, (4) negative predictive value = TU/(TU+FU) = 88%. to check the rate of compliance with record submission (personal information from Table 3: Similarities between true and missed matches. a b AROC). All episodes submitted to AROC were True matches (n=667) Missed matches (n=265) p routinely audited at the point of submission Admitted to major trauma services 631 (95%) 254 (96%) 0.43 (as opposed to regional trauma services) and an audit report of the submitted episodes Age (years, SD) 50 (±22) 48 (±21) 0.15 would have been provided to the facility. Injury Severity Score (SD) 27 (±11) 27 (±11) 0.58 However, the onus was on the facility to make Required intensive care admission 465 (70%) 177 (68%) 0.39 corrections and resubmit where errors were Acute length of stay (days, SD) 40 (±26) 37 (±41) 0.26 found and/or data were missing. Rates of Transferred to specialised rehabilitation 233 (35%) 78 (29%) 0.11 resubmission were not available. (brain injury and spinal injury) a: Trauma records with linked rehabilitation records b: Trauma records with a discharge destination of inpatient rehabilitation but were not linked to rehabilitation records 248 Australian and New Zealand Journal of Public Health 2016 vol . 40 no . 3 © 2015 Public Health Association of Australia Epidemiology Linkage of trauma and rehabilitation datasets This study detected errors in coding in Benefits of this study dataset Acknowledgements discharge destination that affected data The final linked dataset will allow researchers The researchers are grateful to the Motor quality. This could provide an impetus for the to look at the inpatient rehabilitation Accidents Authority of NSW for funding this NSW Trauma Registry to review its coding outcomes of road trauma. It will allow project. We thank CheReL for doing the data system. Both registries could consider having for attempts to develop some models for linkage and the data custodians for the NSW data quality assurance processes in place and estimating functional gains, length of stay in Institute of Trauma and Injury Management could also publish their frequency of checks. rehabilitation, rehabilitation costs based on and AROC for extracting and recoding the An estimate of error from these audits is impairment group and FIM on rehabilitation data for the study dataset. important as a proxy measure of data quality admission. and therefore its usefulness. Although the sensitivity was only 72%, the At all levels, maintenance of a reliable registry References final dataset has 667 valid matches, which such as NSW Trauma Registry is very costly. will make it the largest cohort in the trauma 1. NSW Institute of Trauma and Injury Management. The Linking the NSW Trauma Registry with the 4 NSW Registry Profile of Serious to Critical Injuries, 2008. rehabilitation literature. This study has Sydney (AUST): New South Wales Health; 2011 [cited AROC dataset at various time periods may be demonstrated that the missed links were 2016 Mar 4]. Available from: http://www.aci.health. one option for providing valuable functional nsw.gov.au/__data/assets/pdf_file/0007/195334/ missed randomly. As both datasets were The_NSW_Trauma_Registry_Profile_of_Serious_to_ information for the 29% of multi-trauma population based, the data would provide Critical_Injuries_2008.pdf patients that require inpatient rehabilitation. 2. Dinh MM, Bein KJ, Gabbe BJ, Byrne CM, Petchell J, Lo valid answers to research questions on S, et al. A trauma quality improvement programme As trauma systems in NSW mature and trauma rehabilitation outcomes in NSW. associated with improved patient outcomes: 21 years survival rates increase, this information of experience at an Australian Major Trauma Centre. The process of data concatenation was would be reassuring if one can demonstrate Injury. 2014;45:830-4. considered an important benefit of this 3. Gabbe BJ, Simpson PM, Sutherland AM, Williamson that rehabilitation outcomes are improving OD, Judson R, Kossmann T, et al. Functional measures study dataset. AROC used the following and/or the proportion of patients who are at discharge. Are they useful predictors of longer business rules for concatenation: number of term outcomes for trauma registries? Ann Surg. discharged from rehabilitation with severe days between episodes being 0–14 days for 2008;247:854-9. disability is static or reducing. 4. Khan F, Amatya B, Hoffman K. Systematic review of spinal and 0–7 days for all other impairments multidisciplinary rehabilitation in patients with multiple (personal communication). We felt that these trauma. Br J Surg. 2012;99 Suppl 1:88-96. 5. Simmonds FD, Stevermuer TL. The AROC Annual rules were overly restrictive and, if applied to Conclusion Report: The State of Rehabilitation in Australia in this dataset, would only concatenate 32 sets 2012 [Internet]. Wollongong (AUST ): Australian This study has demonstrated that it is feasible Health Services Research Institute Australasian of patient data rather than 53 sets based on to link road trauma and rehabilitation Rehabilitation Outcomes Centre; 2013 [cited 2016 Mar our clinical reasoning. 4]. Available from: http://ahsri.uow.edu.au/content/ registries using non-identifiable data. groups/public/@web/@chsd/@aroc/documents/doc/ Our paper provided an alternative method However, this then needs to be combined uow160486.pdf for data concatenation for future database 6. Copes WS, Stark MM, Lawnick MM, et al. Linking data with a careful manual review of the results to from national trauma and rehabilitation registries. J studies involving AROC. This study has also make inferences on individual cases. Overall, Trauma. 1996;40(3):428-36. found that, if using concatenated data, the 7. Holman CDJ, Bass AJ, Rosman DL, et al. A decade of this process generated 72% valid matches. real cost of inpatient trauma rehabilitation data linkage in Western Australia: Strategic design, The missed links were missed randomly. applications and benefits of the WA data linkage (based on LOS) would be at least 12% more. This study shows that probabilistic linkage system. Aust Health Rev. 2008;32(4):766-77. 8. Glasson EJ, Hussain R. Linked data: Opportunities and with trauma and rehabilitation datasets is an challenges in disability research. J Intellect Dev Disabil. Future directions option when mapping of the flow of patients 2008;33(4):285-91. 9. Cameron PA, Finch CF, Gabbe BJ, et al. Developing Linkage studies are becoming more prevalent through multiple phases of care in a trauma Australia’s first statewide trauma registry: What are the due to improved methods and are considered system is required. lessons? Aust N Z J Surg. 2004;74(6):424-8. important in the field of injury and disability 10. Da Silveira DP, Artmann E. Accuracy of probabilistic record linkage applied to health databases: Systematic research. However, the quality of linked review. Rev Saude Publica. 2009;43(5):1-7. datasets will depend on the rate of linkage 11. Clarke DE. Practical introduction to record linkage for injury research. Inj Prev. 2004;10:186-91. and the quality of the data used. The rate of linkage could be improved if the NSW Trauma Registry collected unique identifiers. However, if privacy and/or ethical constraints prevent them collecting unique identifiers, they could consider collecting date of birth instead of age. 2016 vol . 40 no . 3 Australian and New Zealand Journal of Public Health 249 © 2015 Public Health Association of Australia

Journal

Australian and New Zealand Journal of Public HealthWiley

Published: Jun 1, 2016

Keywords: ; ; ;

There are no references for this article.