Background Revisions of the International Classification of Diseases (ICD) have previously been shown to cause dramatic effects with regard to injury mortality data when implemented. However, limited knowledge exists on the effects on the coding of external causes of injury morbidity, despite this being an important aspect with regard to injury prevention.
Method Hospitalised injuries in Sweden were studied using time series intervention analysis to observe the effect of the ICD change from ICD-9 to ICD-10 in 1997 on external cause coding.
Results The results would suggest considerable coding issues with a large spike in the proportion of injury admissions registered without an external cause code in 1997, with continuing, although gradually diminishing, problems up to 2002. The coding change seems to have had an immediate effect on all external cause of injury categories, although the categories that were not directly convertible from ICD-9 to ICD-10 were seemingly more greatly affected.
Discussion The study illustrates the potential issues associated with changes between ICD revisions and the importance of data quality control both during surveillance and collection of data, but also when presenting injury trends across ICD versions.
Statistics from Altmetric.com
Injury surveillance is important to understand changes in injury incidence in a defined population over time, to identify vulnerable subgroups within a population and to plan for, and evaluate, injury control strategies. A critical component to such studies is that the collected injury data remain reliable and comparable over time.1
Long-term data collection now exists in many high-income countries and is commonly based on the International Classification of Diseases (ICD) coding system. The ICD is widely recognised as a functional data classification system, categorising injuries by diagnosis (ICD chapter XIX) and by external cause (ICD chapter XX) for morbidity and mortality statistics. In a continuous effort to improve the ICD coding system, periodically, new versions replace older. The current version, ICD-10, was endorsed by the World Health Assembly in 1990 and implemented in Sweden in 1997.
Changes in the ICD coding classification have previously shown to cause dramatic effects on injury mortality data.2–4 While previous studies have shown a continued high quality in hospitalised injury data following the change in ICD coding from ICD-9 to ICD-10,5 ,6 these studies were focused on injury diagnosis rather than the external cause of injuries.
In 2014, the US Department of Health and Human Services issued a rule finalising 1 October 2015 as the compliance date for healthcare providers, health plans and healthcare clearinghouses to transition to ICD-10.7 This decision has been heavily discussed and concerns have been raised of the negative effects of a coding change.8–10
Studying the effect of a coding change on the registration of external causes of injuries is important as injuries are categorised in relation to external causes within different sectors such as transport and workplace or types of injury events such as falls, drowning, violence and so on. Despite the importance of high-quality information on external cause registration for the purpose of injury prevention strategies, previous international studies have highlighted the occurrence of a lack of external cause registration11 and problems in the quality of the registration,12 with regard to injury morbidity. While external cause codes can be estimated based on injury diagnosis, the methods available are not completely reliable and may be prone to error.13 Therefore, to plan and evaluate prevention strategies and to study injury trends in general, a reliable and encompassing registration of external causes is vital.
In Sweden, the validity of external cause registration has previously been shown to be high,14 making the Swedish patient register a suitable register to investigate the impact of the ICD coding change. In the current study, therefore, we aimed to investigate the impact of implementing the tenth revision of the ICD on external cause injury trends in hospitalisations (ICD chapter XX) in Sweden by time series intervention analysis of the proportion of injuries registered with missing external cause codes. We also aimed to categorise any excess injuries with missing external cause associated with the transition by analysing trends in the incidence of hospital admissions due to injuries coded within broad external cause categories.
Materials and methods
Sweden has, through the National Patient Register (NPR), collected data on hospital admissions due to injuries since 1964, but it was not until 1987 that the reporting system was all-encompassing, covering all Swedish counties, and considered reliable.14 Since then, Sweden used two different versions of the ICD system. ICD-9 was used until 1996 and ICD-10 from 1997 and onwards.
Data from the NPR were compiled regarding hospitalised injuries both with and without external cause, with information on gender and age, on a yearly basis, from 1987 to 2012. To ensure comparability between versions, external causes were categorised in broad categories based on the possibility of converting ICD-9 codes to ICD-10 codes. The broad categories were self-inflicted injuries (950–958 (ICD-9), X60–X84 (ICD-10)), falls (880–888 (ICD-9), W00–W19 (ICD-10)), undetermined intent (980–988 (ICD-9), Y10–Y34 (ICD-10)), transport (807–849 (ICD-9), V01–V99 (ICD-10)), assault (960–968 (ICD-9), X85-Y09 (ICD-10)) and miscellaneous (all external cause codes within the interval 800–999 (ICD-9) and V01–Y98 (ICD-10) not included in the other categories). While it had theoretically been possible to convert more ICD-9 codes to ICD-10 codes in the miscellaneous category, this would have required the use of more specific ICD-9 codes. Due to a lack of the necessary specific coding beyond the first digit of external cause codes in some cases, this was unfortunately not possible. In proportional terms, the miscellaneous category consists predominantly of codes related to misadventures to patients during surgical and medical care along with later complications of medical care, followed by exposure to both animate and inanimate mechanical forces.
The data used in this paper are generally available and anonymised from the National Board of Health and Welfare, therefore exempting the need for ethical consideration.
Time series analysis of the proportion of admissions with missing external cause code (of all admissions due to injuries) was performed using a generalised additive model for location, scale and shape (GAMLSS) assuming a β distribution.15 ,16 The log-link was used to obtain rate ratio (RR) estimates. Dummy variables were created to reflect each successive year after the coding change to test whether they deviated from the underlying secular trend, which was estimated using cubic smoothing splines.17 Residual autocorrelation was checked for using a robustified portmanteau test with automatic lag selection.18 The test statistic was not significant, indicating that the residuals were free of autocorrelation. Estimations of percentage point differences from the secular trend and the estimated number of additional admissions with missing external cause code attributable to the coding change were calculated by taking the difference between the observed values and the predicted values for the secular trend from the model. Separate univariate autoregressive moving average (ARMA) models were then employed to hospital admissions data by external cause category to estimate the origin of the excess missing external cause codes. Linear trend and dummy variables for the first 5 years after the transition to ICD-10 were included in these models. The optimal ARMA order for each time series was chosen based on studying plots of the autocorrelation function (ACF) and partial ACF,19 along with an iterative optimisation process based on minimising the Akaike information criterion, that is, choosing the autoregressive (AR) or moving average (MA) parameters that provided the best fit. Only significant AR or MA parameters were retained in the models. There was no evidence of residual autocorrelation, and a Jarque–Bera test20 showed no evidence of non-normality in the residuals in any of the Gaussian models. Adjusted trend lines for the period 1997–2001 were calculated by taking the inverse of the parameter estimates from the model and adding the result to the fitted values (thus removing the estimated intervention effect). The statistical analysis was performed in R using the GAMLSS, vrtest, tseries and forecast packages.21–25
Between 1987 and 2012, the number of injuries registered in the Swedish NPR was 3 776 867. During the same time period, the number of injuries registered with no external cause was 135 010, that is, on average 3.6%. However, as is seen in figure 1, the distribution of these was highly uneven over the course of the study period. Three separate time periods can be identified. First, an ICD-9 period where the proportion of injuries registered with missing external cause was kept at a reasonably low level (2%–4%) between 1987 and 1996. Second, at the time of the transition from ICD-9 to ICD-10, the proportion of injuries registered with no external cause increased to 18.5% and did not return to under 5% until 2001. During the third time period (between 2002 and 2012), the annual proportion of missing external causes was generally below 2.5%.
Due to the third time period of low annual proportions, the time series analysis showed a decreasing trend towards the end of the study period. Results from the generalised additive model also showed statistically significant deviations from the predicted underlying secular trend between the years 1997 and 2001, with RR estimates ranging from 6.18 in 1997 to 1.41 in 2001. In total, the estimated excess of injuries without an external cause code was estimated at 47 773 for the 5-year period after the transition, which is an increase of 243% compared with the predicted trend (table 1).
As is seen in table 2, the registration of all investigated categories of external causes was significantly lower than expected during the first year of ICD-10 (1997), with the greatest percentage decrease among injury admissions with undetermined intent (−49.2%) and the greatest absolute decrease in the miscellaneous external cause category (−11 812). In the second year after the coding change (1998), admissions were still significantly lower than expected in all external cause categories apart from falls and transport-related injuries, which appear to have returned to normal after 1 year. In most cases, with the exception of injuries with undetermined intent, the difference between the observed and expected values was smaller in comparison with the first year. In the third year after the transition to ICD-10 (1999), only self-inflicted injuries, injuries with undetermined intent and miscellaneous injuries showed significant deviations from the underlying trends. Self-inflicted injuries was the only category that deviated significantly during the fourth year (2000), and by the fifth year (2001) there were no statistically significant deviations in any of the external cause categories. Adjusted trend estimations (with 95% CIs) during the first 5 years after the coding change for all investigated external cause categories are displayed in figure 2.
Counting only the statistically significant estimates from the ARMA models and relating them to the results from the proportional GAMLSS model, 50% of the excess injuries with missing external cause during the period 1997–2001 (n=47773) are estimated to belong to the miscellaneous external cause category, and 19%, 9%, 5%, 3% and 2% are, respectively, estimated to belong to the falls, self-inflicted, undetermined intent, transport and assault categories, while 12% are of unknown origin.
This study can show that the changes in ICD coding between ICD-9 and ICD-10 seems to have caused problems in the registration of hospitalised injuries in Sweden. The proportion of injuries with missing external cause were kept at a low level throughout the ICD-9 period to then sharply increase in 1997, with low levels not being reassumed until 2002.
While only broad categories of external causes could be used due to conversion issues, the results provide further insight as to the impact of the coding change. All investigated external cause categories seem to have been affected during 1997, with the greatest relative impact observed in injuries with undetermined intent. The greatest absolute impact was seen in the miscellaneous category that consists mainly of injuries associated with healthcare services. While it thus stands to reason that the majority of the absolute effect may be related to such injuries, the exact extent to which each of the external causes included in this very broad category were affected relative to their own trend remains unknown.
The fact that an immediate as well as a prolonged effect of the ICD coding change was observed both with regard to the missing external cause data and the difference in effect on other external causes indicates two different effects of the coding change. The immediate effect during the first year after the coding change seems to have affected all types of injuries, regardless of cause. This effect would indicate an issue directly connected to the process of changing from a known coding system to an unknown coding system. A crucial difference between the two coding systems is the number of codes available with almost five times as many diagnosis codes in ICD-10 compared with ICD-9.
We also found evidence of a prolonged effect (1998–2001) following the coding change in which injuries with missing external cause were found to be significantly higher than expected. This effect was consistently lessened for each year and seems to be more connected to certain types of external causes. Specifically, the prolonged effect seems to highlight another issue with the transition from ICD-9 to ICD-10: the problems with compatibility and conversion between the two coding systems. While some external code categories are convertible from the primary code, for example, fall-related and transport-related injuries, for many other external causes this is not possible. The results in this study seem to suggest that this aspect of the coding change may have prolonged the period of lower-than-average data quality in relation to external cause coding. This is most clearly seen in the lingering effect on the miscellaneous category, while the more easily convertible codes (fall-related and transport-related injuries) returned to the estimated level within 1 year. While none of the subcategories within the miscellaneous category are among the leading causes of injury morbidity in the population as a whole, having reliable long-term data are nonetheless crucial in the injury prevention community and there is a considerable risk that trend data across two ICD periods can be distorted.
Although not shown in the results, we also studied the missing external cause data in age-specific and sex-specific groups. However, no decisive patterns were found, with similar results being produced regardless of age or sex. Therefore, this would indicate that although certain types of external cause codes were affected to a greater degree, the population was affected in a similar fashion.
While this study highlights important issues and problems with a change in classification systems, it also indicates that using ICD-10 produces an improved quality of external cause data than ICD-9. When comparing the percentage of injuries registered without an external cause during the period 1987–1996 (ICD-9) and 2002–2012 (ICD-10), a considerably lower number of injuries were registered without an external cause during the later time period, with between 2% and 4% during the first period and below 2.5% during the last period. Although this positive effect is not necessarily due to the change from ICD-9 to ICD-10, it would seem likely that the increase in diagnosis codes improves the registration quality, once physicians and registrars have become familiar with the new system. This is also illustrated by the fact that the registration of injuries as a result of surgical and medical care increased during the ICD-10 period, within the miscellaneous category. According to The National Board of Health and Welfare, this was not due to specific government guidelines, rather an effect of the increase in possible external injury codes (Pernilla Fagerström, personal communication, 20 August 2014).
The results shown in this study, we would argue, are highly relevant from both a national and international perspective. First, Swedish morbidity data are recognised as a reliable, long-term data source. With potentially unreliable periods of data, this reputation is considerably affected and raises questions concerning the previously published historical trends of hospitalised injuries based on chapter XX. Second, a planned ICD coding change from ICD-9 to ICD-10 is planned in the US during 2015 and these results could be crucial in minimising the impact on data quality in relation to external causes of injuries. Third, a change from ICD-10 to ICD-11 is under way26 and it may be of benefit to include the findings in this paper in the discussions concerning the national implementations of the 11th version of the ICD. If similar problems occur at the next ICD change, presenting reliable historical injury trends and future prognoses will be difficult unless attempts at statistical adjustments to account for this are made (which inherently are susceptible to some error). Although a sudden and large outlier, such as the effect in 1997 may be found when analysing other external cause of injury trends, the smaller, but lingering, problems such as the ones observed between 1998 and 2001 may not necessarily be discovered without prior knowledge of such issues. Finally, while this study has shown quasi-experimental evidence of a negative effect on injury morbidity data quality in Sweden, it would be unwise to suggest that these problems are limited to a particular country. Rather, it may be expected that similar issues could be found elsewhere.
There are some limitations to this study that are important to acknowledge. First, although sound statistical methods have been used, the results in this study are estimates and do not necessarily represent a true picture of the data. The estimates of where the injuries with missing external cause originated, for example, is based on the trends before and after the coding change. Although it is likely that similar trends were seen during 1997–2001, this is an assumption that is not necessarily true and that could affect the results. Second, while a majority of the injuries with missing external cause could be accounted for, approximately 12% of the injuries with missing external cause could not be categorised. For instance, none of the excess injuries with missing external cause in the fifth year after the transition to ICD-10 found by the proportional (GAMLSS) model could be categorised. This could be due to a lack of statistical power to detect an effect of a smaller size compared with the earlier years post-transition in the individual external cause incidence (ARMA) models. Further attempts at improving the precision of the statistical models to categorise these could potentially affect the results though we believe the effect on the conclusions drawn from this study would be minimal. Finally, during the period 1997–2001 there may have been other interfering events that would confound the results apart from the coding change. To our knowledge no other major changes (eg, considerable changes in the healthcare system) occurred during this time period.
One of the goals of injury registration is to present historically reliable injury data that can be used as a foundation for healthcare planning. Researchers, government officials and practitioners alike need to be aware of the risks in changing ICD versions and balance these against the continuing need for improvement. Given the importance and key role of injury surveillance in the historical injury prevention work, it is vital to address the presented issue to ensure that similar situations do not arise in the future.
$150 million award against Jeep
A jury in Georgia awarded $150 million to the family of a 4-year-old boy killed when a Jeep Grand Cherokee exploded into flames after being rear-ended. Two years ago Chrysler agreed to a limited recall of some older-model Jeeps with the rear-mounted tanks. About 75 people have died in postcrash fires because of the fuel tank design. (Noted by IBP)
What is already known on this subject
Reliable injury data are vital for successful injury prevention strategies.
Problems in the quality of injury data have previously been seen between International Classification of Diseases (ICD) versions.
What this study adds
In conjunction with the transition from International Classification of Diseases (ICD)-9 to ICD-10, the percentage of hospitalised injuries registered with missing external cause increased considerably.
Elevated levels of injuries with missing external cause code continued for 5 years after the ICD coding change.
Registration problems seem to have predominantly occurred within injury categories where direct conversions between primary ICD-9 and ICD-10 codes were absent.
Contributors FN conceptualised and designed the study, carried out the analysis, drafted the initial manuscript and approved the final manuscript as submitted. CB assisted in the design of the study, carried out the analysis, reviewed and revised the manuscript and approved the final submitted manuscript. RA assisted in the design of the study, reviewed and revised the manuscript and approved the final submitted manuscript.
Competing interests None.
Provenance and peer review Not commissioned; externally peer reviewed.
Data sharing statement All data included in this study are obtained through the Swedish hospital discharge register. These data are generally available from the Swedish National Board of Health and Welfare.
If you wish to reuse any or all of this article please use the link below which will take you to the Copyright Clearance Center’s RightsLink service. You will be able to get a quick price and instant permission to reuse the content in many different ways.