Developing the ICD-10-CM crosswalk and definitions was an iterative process. First, we adapted the set of broad 57 ICD-9-CM ClinDx groups (Meyers et al., 2018) to create a concordance table, similar to the broad, aggregate ICD mortality groups that Janssen and Kunst (2004) used to evaluate mortality trends across ICD versions. In short, we identified the best match for each ICD-9-CM ClinDx group among the ICD-10-CM chapter(s) (such as Neoplasms), categories or sub-categories. Each ClinDx group name was edited to improve consistency and align with ICD-10-CM descriptions. After developing the ICD-10-CM ClinDx groups, we aligned ICD-9-CM codes with the updated ClinDx group case definitions. Both backward and forward mapping techniques were used to improve the accuracy of the mappings to each ClinDx group. This strategy reduced challenges with one-to-many or many-to-one mappings. For example, “Hernia of abdominal cavity” (550–553) aligned with the “Hernia” category (K40-K46). The new Hernia ClinDx group retains the title consistent with ICD-10-CM terminology and includes all hernias. Then, all codes from the Hernia of abdominal cavity ICD-9-CM ClinDx group were assigned to the new Hernia ClinDx group.

This study used ICD-9-CM codes within the range of 000–999 and ICD-10-CM codes beginning with letter A–U (2021 codes and prior years). Please note, whenever a specific diagnosis code is mentioned in this paper, if the first character is numeric, it is an ICD-9-CM code, whereas ICD-10-CM codes begin with a letter. External cause and history codes were excluded. Quality checks of draft crosswalks were conducted in SAS^® (Release 9.4, SAS Institute, Inc., Cary, North Carolina) and Microsoft Excel by two investigators, a Research Epidemiologist and a Registered Health Information Administrator (RHIA). A physician/medical epidemiologist was consulted as needed. In some cases, discontinuities identified during initial interrupted time series analyses (Section 3.3) identified errors that were subsequently corrected in the final version. For example, to align with ICD-10-CM, ICD-9-CM tooth fracture codes were moved to Fracture – head ClinDx group from the Open wounds nec group. Corrections were made in SAS^®, and the process was repeated until both reviewers were satisfied with the accuracy of the ICD-10-CM crosswalk. All records were assigned to a single ClinDx group using a SAS^® program that included several elements of the Barell Matrix SAS programs (National Center for Health Statistics, 2002).

The Ohio Bureau of Workers’ Compensation (OHBWC) is unique because OHBWC insures about two-thirds of workers in Ohio. Ohio is the largest of four states where it is mandatory to maintain WC insurance from the state-based insurer, with the exception of some employers who meet the financial criteria to self-insure and prefer to self-insure (Ohio Bureau of Workers’ Compensation, 2024) and sole proprietors, who are not required to insure through OHBWC.

NIOSH maintains a large relational SQL database that includes variables provided by OHBWC that were used to describe the characteristics of the study population of diagnoses and claims overall and stratified by transition stage (pre-transition, 1/1/2001–9/30/2015; post-transition, 10/1/2015–12/31/2018). For claims data from 2011 to 2018, we counted distinct ClinDx groups per claim among all Ohio employers insured by the OHBWC. That is, we counted each group once per claim even when there were multiple diagnoses assigned to the same ClinDx group. For this study, analyses excluded 2,148 (0.32%) accepted claims due to: zero diagnosis codes; diagnosis denial; the injured worker requested dismissal of the condition; black lung diagnoses, which are handled in a separate system; incomplete diagnosis codes; or codes mapped to multiple ClinDx groups.

This activity was reviewed by CDC, deemed not to be research, and was conducted consistent with applicable federal law and CDC policy (See e.g., 45 C.F.R. part 46.102(1)(2), 21 C.F.R. part 56; 42 U.S.C. §241 (d); 5 U.S.C. §552a; 44 U.S.C. §3501 et seq.).

Interrupted time series analysis models were used to estimate the effect of the transition on the monthly distribution by distinct ClinDx groups (%). The main purpose of the regression analysis was to identify changes in level (monthly percentage) or slope (monthly trend). A level change is a sudden, sustained change where the time series is shifted up or down by a given value immediately following the transition. For regression analysis of distinct ClinDx groups, the denominator for percentage calculations was the sum of distinct ClinDx groups per claim among all claims. A separate model was used for each of the 57 ClinDx groups. The model used was: Yt=β0+β1month+β2transition+β3month–57×transition+εt, where Yt is the monthly ClinDx group percentages measured at each month of observation t, β0 is the intercept, the baseline outcome proportion in January 2011, β1 is the slope or trend over months, β2 is the level change at the transition, β3 is the change in slope at the transition, and errors are εt. The variable month ranges from 0 to 95 for the months January 2011 to December 2018. Month 57 was the month of transition. The variable transition was coded as zero if the month occurred before October 2015 or one if the month occurred on or after October 2015. To account for seasonal effects, we added eleven variables to the model to indicate the months February to December. January was the reference month. An autoregressive model with lags up to 24 months was used for the errors, εt. All autoregressive parameters were initially included in the model, and parameters were removed by backward elimination.

Ten of the ClinDx groups had a low number of counts and no visual evidence of a trend or seasonal effect; the model was simplified to: Yt=β0+β2 transition. A negative binomial model for counts was used to fit these data, with an offset to account for the total number of diagnoses in each month.

Relative level changes (%) compared to September 2015 were calculated to normalize comparisons between ClinDx groups despite large monthly frequency differences. For linear models, the percent change at transition was equal to 100 times the ratio of the level change coefficient (β1) to the difference between the post-transition intercept (β2) and the level change coefficient [100*(β1/β2-β1)]. For negative binomial models the percent change at transition was equal to 100 times the difference between the exponentiated coefficient minus 1.

All calculations were done with SAS^®. The AUTOREG and COUNTREG procedures were used to estimate the models. The residuals of the models were evaluated with the ARIMA procedure and were found to be white noise.

Descriptive analyses for this study were reported overall and stratified by transition stage. We calculated the distribution stratified by diagnosis type (injury or disease) for each claim, for diagnoses that met our inclusion criteria, and for each distinct ClinDx group. For each claim, we calculated the number and percentage of claims that included at least one injury and disease diagnosis, included a single diagnosis code, or included a single ClinDx group. We calculated the mean and standard deviation (SD) overall and by transition stage for the numbers of allowed diagnoses per claim and distinct ClinDx groups per claim.

For ClinDx groups with statistically significant level changes of ≥ 20% at the transition, we reviewed diagnosis frequencies to identify potential errors in the code assignments that could be corrected or to identify other potential reasons for the level changes. After correcting errors, models were rerun.

The final set of 57 mutually exclusive ICD-10-CM ClinDx groups (Supplementary data Table 1) was similar to the 57 ICD-9-CM groups (Meyers et al., 2018) with four main exceptions. First, to improve tracking capabilities for knee injuries made possible by the ICD-10-CM coding structure, we created a new group — the Knee sprains or tears group — which includes knee meniscus tears (previously classified as knee dislocations and knee ligament sprains). ICD-9-CM codes for “unspecific sprains of the lower leg and knee” remained in the Sprains – lower extremity group. Second, we combined knee derangements with other knee joint disorders into a revised group, Knee disorders. Third, we eliminated the Certain traumatic complications and unspecified injuries (ICD-9-CM 958.0–959.9) ClinDx group and assigned those codes to the Other and unspecified effects of external cause not elsewhere classified (nec) group. Fourth, the newborn codes (Perinatal codes, ICD-9-CM 760–779) were assigned to Congenital abnormalities nec and perinatal conditions, rather than Pregnancy related conditions and adverse outcomes.

Minor changes included editing the names of 11 groups to align with ICD-10-CM terminology and moving 50 ICD-9-CM codes from four ClinDx groups to different groups (34 of the 50 diagnoses did not exist in the OHBWC claims database, and 12 diagnoses had < 50 claims). ClinDx group descriptions, case definitions, and ranges of ICD-9-CM and ICD-10-CM codes per group are presented in Supplementary data Table 1.

When stratified by transition stage, patterns in the percentage of allowed diagnoses, in distinct ClinDx groups, and in distinct claims were not meaningfully different for any category (Tables 1 and 2). Across both time periods, among 661,683 claims, more than 61% had a single diagnosis (Table 1), about 70% included diagnoses from a single ClinDx group, and about 96% included at least one injury diagnosis. These claims included 1,171,773 diagnosis codes (Table 2), primarily for traumatic injuries (90%). The mean (SD) allowed diagnoses per claim was 1.77 (1.50) overall, 1.78 (1.51) pre-transition, and 1.75 (1.47) post-transition (data not included in Tables).

The five most common ClinDx groups were the same for all time periods (Table 3): 17.96% Open wounds nec, 16.62% Contusions with intact skin surface (Contusions), 10.11% Sprains – upper extremity.

(Sprains-UE), 9.91% Sprains – back (Sprains-back), and 7.25% Sprains – lower extremity except knee (Sprains-LE). The order for third and fourth ranked groups switched post-transition because there were fewer back sprains. Seventeen ClinDx groups with > 12,000 claims accounted for 90% of all ClinDx groups (N = 864,419); the 17 groups were primarily traumatic injuries and 2 disease groups – Soft tissue/enthesopathy ranked ninth (N = 26,656, 2.78% of ClinDx groups, 4.03% of claims) and Diseases of the nervous system and sense organs nec ranked 17th (N = 12,197, 1.27% of ClinDx groups, 1.84% of claims). When combined, the six ClinDx groups for strain and sprain diagnoses by body region accounted for about 35% of distinct ClinDx groups. The only group with zero claims was Death, cause unknown.

Based on a frequency analysis of the most common diagnoses by ClinDx group, it seems the case definitions for 13 of 17 ClinDx groups with statistically significant level changes (N = 4 large level changes) were correctly classified and lacked a clear explanation for the level change other than the change in coding systems. The thirteen groups were: Amputation, Carpal tunnel syndrome, Contusions with intact skin surface, Diseases of the digestive systems nec, Diseases of the skin and subcutaneous tissue nec, Dislocation, Fracture – lower extremity, Infectious and parasitic diseases, Other and unspecified effects of external cause nec, Soft tissue/enthesopathy, Sprains-back, Sprains – nec, and Superficial injury. For the other four ClinDx groups with statistically significant level changes, we identified three reasons for the level changes: (1) increased specificity for some lower extremity sprain and strain codes; (2) some external cause codes changed to diagnosis codes; or (3) enactment of a firefighter cancer presumption law. First, level changes for two groups, Knee sprain or tear and Sprains-LE, were caused by two ICD-9-CM codes for Sprains and strains of other specified (844.8) or unspecified (844.9) site(s) of “knee and leg.” Although we assigned 844.8 and 844.9 to the Sprains-LE group, a priori, some portion of those relatively common diagnoses was miscategorized and should have been in the Knee sprain or tear ClinDx group. Therefore, pre-transition Knee sprain or tear group results were underestimated, and pre-transition Sprains-LE were overestimated. Second, because ICD-9-CM identified some poisonings using external cause codes, this may account for the relatively large increase in Poisoning and toxic effects levels (Annest, Hedegaard, Chen, Warner, & Small, 2014; Slavova et al., 2018). Lastly, the large increase in Neoplasms at the transition may be associated with an increase in neoplasm claims from firefighters due to the Ohio firefighter cancer presumption law that became effective in April 2017 (Ohio Legislature 135th General Assembly, 2017). It is unclear whether the timing of the new law explains the immediate level increase in October 2015; however, preliminary analyses (data not shown) suggest that the percentage of neoplasms attributable to firefighters increased substantially post-transition, at a time when mesothelioma claims were declining. October 2015 claims could have been filed well after the transition because the Ohio statute of limitations for filing a disease WC claim is two years.

Based on an examination of specific diagnoses codes in the Sprains-nec group, it appeared that an increased number of torso region strains and sprains may be contributing to the post-transition changes; these included thorax strain codes (S29.011A, S29.012A, S29.019A); thorax sprain codes (S23.8XXA, S23.9XXA); and an abdominal strain code (S39.011A). We cannot determine whether torso region strains and sprains have increased post-transition because those diagnoses would have been coded as Other specified sites of sprains and strains (848.8); a more specific diagnosis code was unavailable. In terms of the Sprains-nec slope change, perhaps some ICD-10-CM diagnoses in Sprains-nec are being used to replace less specific ICD-9-CM diagnoses in Sprains-UE. If so, that could explain the change to a positive slope post-transition. However, we were unable to identify specific codes that would support that theory.

Despite the major changes introduced by the transition, we successfully created a manageable set of 57 broad ClinDx groups optimized for occupational health surveillance using all ICD-9-CM and ICD-10-CM injury and disease diagnosis codes. While the number of diagnosis codes increased five-fold with the transition from ICD to 9-CM to ICD-10-CM, in this study the mean number of diagnoses and distinct ClinDx groups per claim remained consistent during each transition stage. Pre- and post-transition, the rankings of the top five distinct ClinDx groups included the same groups, with a slightly different order. In this large study, while diagnoses were predominantly injury codes, we counted at least one disease diagnosis for 11.77% of claims. These findings underscore the need for a single system to characterize both injury and disease outcomes when analyzing occupational health data. Specifically, among disease outcomes, diagnoses from the Soft-tissue/enthesopathy musculoskeletal diseases and Diseases of the nervous system and sense organs nec ClinDx groups were the most common; those two groups were unaffected by the transition in terms of slope changes, and the level change for Soft-tissue/enthesopathy was relatively small (−10.8%).

It is our hope that this crosswalk will benefit projects that use administrative billing data to track or analyze work-related injuries and diseases that bridge the transition. The revised crosswalk is improved and corrected some minor errors in the prior version (Meyers et al., 2018). The comprehensive list of ClinDx groups by ICD-9-CM and ICD-10-CM codes (and header rows) will be updated regularly; for the most current listing (one row per code), contact the NIOSH Center for Workers’ Compensation Studies at cwcs@cdc.gov.

To the best of our knowledge, these 57 ClinDx groups represent the only classification system with both disease and injury diagnoses using ICD-9-CM and ICD-10-CM codes. This feature is important because large administrative data sets such as the OHBWC claims data include records for both versions during an injured worker’s medical journey. Studies with small study populations would be underpowered if they needed to exclude pre-transition diagnosis data due to the transition. To maximize the utility of administrative data sets like these OHBWC data, it is vital to take advantage of all available data, especially when analyzing high-risk occupations and industries with relatively small numbers of policies, employees, and claims.

In the literature, three strategies recommended for overcoming inconsistencies introduced by the transition to a new ICD revision include: (1) using broad groups of diagnosis codes for the same medical concept (such as ICD chapters); (2) using concordance tables based on high-level ICD categories (such as three-digit codes, four-digit codes); or (3) analyzing specific diagnoses where consistency in coding is expected across the transition (Alter & Carmichael, 1996; Janssen & Kunst, 2004; Pickett, 2014). All three strategies were used to create our 57 ClinDx groups.

For the 26 traumatic injury ICD-10-CM chapter ClinDx groups, our code assignments were consistent with the ICD-10-CM injury diagnosis matrix (Hedegaard, Johnson, Garnett, & Thomas, 2020). The few minor differences we identified were for rarely used codes and would not have affected our results.

The six ClinDx groups used to classify strains and sprains accounted for 35% of distinct ClinDx groups. In some cases, it may be beneficial for investigators to create additional body-region-specific ClinDx groups for strains and sprains or use the ICD-10-CM injury diagnosis matrix body regions (Hedegaard et al., 2020). The Sprains-nec group mostly includes strains and sprains from the torso (chest, abdomen, pelvis) and the head (such as jaw sprains).

In this large population of WC claims from Ohio, the effect of the transition varied by ClinDx group. We observed statistically significant level or slope changes for 22 ClinDx groups. Specific reasons for the discontinuities were unclear, with four exceptions, as noted in section 3.4. Other studies have also found discontinuity in Poisonings and toxic effects, which they attributed to reclassifying some poisonings from external cause codes to diagnosis codes (Annest et al., 2014; Slavova et al., 2018). For this study, when interpretating the effect of the transition on the slope changes by ClinDx group we were unable to find alternative reasons in society that would have impacted WC claim slopes or workplace hazards immediately after the transition. Administrative medical coding changes can affect records associated with WC claims but has zero effect on work-related injuries/disease occurrence or WC claim filing patterns. Other investigators have noted situations where there was an adjustment period early in the ICD-10-CM era where health systems adjusted to the new coding rules and new, more specific ICD-10-CM codes that were not represented in ICD-9-CM (Ellis et al., 2020; Salemi, Tanner, Kirby, & Cragan, 2019; Slavova et al., 2018). This could account for some relatively small changes observed at the transition where the levels in 2018 were similar to the levels just prior to the transition. However, typically if you extended the pre-transition trend lines to December 2018, the lines would not intersect with the post-transition trend. ICD-9-CM traumatic injury codes changed more dramatically than any other ICD chapter, with a 15 -fold increase in the number of codes and over half with convoluted mappings (where “multiple ICD-9-CM and ICD-10-CM codes share complex, entangled, and non-reciprocal meanings”) or no mappings (Boyd et al., 2013). This may have contributed to some of the discontinuities caused by the transition, where specific reasons were unclear.

Existing research on the effect of transitioning between ICD versions is not directly comparable to this study using WC data. Other studies have analyzed data from the military or general healthcare settings by encounter (De Crescenzo, Gabella, & Johnson, 2021; Duchesneau, et al., 2023; Ellis et al., 2020; Gibson, Casto, Young, Karnell, & Coenen, 2016; Hirai, Owens, Reid, Vladutiu, & Main, 2022; Hsu et al., 2021; Inscore, Gonzales, Rennix, & Jones, 2018; Mayhew et al., 2019; Salemi et al., 2019; Sanusi et al., 2022; Schuh-Renner, Inscore, Hauschild, Jones, & Canham-Chervak, 2021; Sebastião, Metzger, Chisolm, Xiang, & Cooper, 2021), rather than our study population of non-military, work-related outcomes where claims-level data have had about two years to mature. Among other studies, some investigators have also conducted segmented regression analyses to evaluate level and slope changes for specific diagnoses or groups of diagnoses (De Crescenzo et al., 2021; Duchesneau et al., 2023; Ellis et al., 2020; Hirai et al., 2022; Hsu et al., 2021; Sebastião et al., 2021). Among these, our study design most closely resembled the comprehensive analyses of monthly rates in the IBM MarketScan Commercial Database (N = 3.4 billion diagnoses) using three different comprehensive classification systems; one system they evaluated was the World Health Organization (WHO) disease chapters, which align with eight of our ClinDx groups (Ellis et al., 2020). Their analysis of WHO disease chapters found zero statistically significant level or slope changes greater than 20%. In contrast, among the eight ClinDx groups equivalent to WHO chapters, three relatively small groups had large level or slope changes at the transition (large level changes: Infectious and parasitic diseases, Neoplasms; large slope changes: Symptoms, signs, abnormal clinical or laboratory findings nec). However, that study used a conservative Bonferroni correction, which required P < 0.0027 to be considered statistically significant. We could not compare our results to uncorrected significance levels because those data are unavailable. Also, the claims data for this study were dated by injury date and thus included some ICD-10-CM codes in the pre-transition stage; whereas, Ellis and colleagues (2020) analyzed data by encounter date.

When working with diagnosis data that bridge the October 2015 transition, we propose three main strategies for overcoming the effect of the transition — adjust, combine, or exclude (ACE) (Gibson et al., 2016; Janssen & Kunst, 2004); that is, either adjust pre-transition distributions, combine ClinDx groups, or exclude some ClinDx groups or all pre-transition data. Even if investigators decided to analyze pre- and post-transition data separately, using the most current version of the ClinDx crosswalk is recommended because it aligns with ICD-10-CM.

For the three ClinDx groups where clear, major changes in ICD coding explained statistically significant level and slope changes, all three strategies could be used to overcome the problem. For example, to correct overestimates for Sprains-LE and underestimates of Knee sprains and tears group, one could exclude pre-transition data for the two interrelated groups. Alternatively, combining the Sprains-LE with the Knee sprains and tears could work when including both pre-transition data is preferred. However, the Knee sprains and tears group was common, and keeping the groups separate could identify more specific contributing factors. Another option would be to use a comparability analysis to adjust pre-transition levels for both ClinDx groups (Pickett, 2014). For example, rather than assigning all 844.9 claims to the Sprains LE ClinDx group, those claims could be redistributed to the Sprains LE and Knee sprains and tears ClinDx groups based on the distribution of all ICD-10-CM diagnoses for unspecified knee sprains and unspecified lower leg sprains. Depending on the project goals and size of the study population, pre-transition Poisoning and toxic effects level underestimates could be adjusted or excluded when studying this ClinDx group across the transition period.

More work is needed to explore how to use ITS coefficients to adjust pre-transition percentages to estimate an ‘expected’ number of distinct ClinDx groups that would have been coded to a particular group post-transition. A similar approach was used in the UK after the transition to ICD-10 mortality codes (Brock, Griffiths, & Rooney, 2006; Griffiths, Brock, & Rooney, 2004). The UK Office of National Statistics analyzed changes to respiratory or circulatory disease mortality code assignments in ICD-10 to estimate how the transition affected disease deaths rates in England and Wales. Those reports presented unadjusted and adjusted trends after adjusting some ICD-9 mortality estimates using comparability ratios calculated by age and sex. However, comparability ratios (also called comparability factors) typically use dually coded data to calculate the ratios (Gibson, et al., 2016). For example, for future work using these OHBWC claims data, quantitative results from ITS models could be used to adjust pre-transition distributions or slopes. Depending on a project’s purpose, it may be appropriate to combine uncommon groups from the same ICD chapter (such as combining all congenital diseases, combining all circulatory system diseases) or to create residual categories for all “uncommon” injury or disease ClinDx groups below a specified percentage threshold. However, due to the highly skewed nature of WC claim costs (Chadarevian & Myers, 2023), we caution against selecting arbitrary frequency thresholds before considering the influence of high-claim cost outliers. While we have suggested using ACE strategies, in some situations (such as projects that avoid time series analyses) investigators may disregard the effect of the transition on their results and acknowledge the resulting misclassification as a limitation.

The main strength of this study was the large number of claims, diagnoses, and ClinDx groups included in the analysis. This study included almost five years of pre-transition data and three years of post-transition data and included all compensable diagnoses for approximately two-thirds of workers in Ohio during the eight years. Also, using an exhaustive and mutually exclusive set of 57 ClinDx groups was a strength because a change for one group was likely offset by an equal and opposite change in another group. For example, similar level change values in opposite directions for Sprains-nec and Sprains-back, 0.482 and −0.442%, may be related. In contrast, most other studies have focused on evaluating the effect of the transition on a subset of specific ICD-CM codes or diagnosis types, such as birth defects (Salemi et al., 2019), common conditions (Hsu et al., 2021), common pain conditions (Mayhew et al., 2019), traumatic brain injury (De Crescenzo et al., 2021), or external-cause-of-injury (Slavova et al., 2018).