The codes of the International Classification of Diseases (ICD) serve as the backbone for billing, payment, and surveillance programs across the entire healthcare system – nationally and globally. Recent research published in Medical Care by Alexander Mainor and colleagues from the Dartmouth Institute for Health Policy & Clinical Practice has shown that the transition of the ICD from version 9 to version 10 may have caused more turbulence than previously acknowledged. To be clear – this is not just a nerdy curiosity of measurement methodology. This is the kind of thing that leads people to question statistical reports in this “post-truth” era.
A little history
Standardized coding medical conditions may have started with John Graunt’s statistical studies of the London ‘bills of mortality’, or possibly with Francois Bossier de la Croix’s 18th-century text, Nosologia methodica. Don’t worry, I’m kidding! That is probably a little too much for now . . . but history nerds will definitely want to check out this historiography of the International Classification of Diseases by the World Health Organization (WHO) [pdf].
Let’s fast forward: Since 1979, the US healthcare system operates on the backbone of a set of codes, called the International Classification of Diseases (ICD) version 9. Although most of the world has adopted this standard championed by the WHO, the National Center for Health Statistics (NCHS) and the Centers for Medicare and Medicaid Services (CMS) jointly act as the primary agencies maintaining and overseeing these codes for our country. The US’s clinical modifications to ICD-9 are referred to as the ICD-9-CM.
These codes are used to encapsulate and standardize medical diagnoses made and, of course, billing. While there are procedural codes in both ICD-9 and -10, procedure billing still heavily relies on a separate set of codes labeled Current Procedural Terminology (CPT), organized by the American Medical Association. Given the national and even global importance inherent in updating the US healthcare diagnostic and procedural codes, the process is – to put it mildly – pretty darn complicated [pdf].
Why was there a transition from ICD-9 to ICD-10?
There were many reasons why replacing the ICD-9-CM codes was a good idea. For one, the ICD-9-CM tabular list was running out of numbers, and new code proposals were being denied due to space limitations. Many people complained that the ICD-9 codes and V-codes, in particular, lacked sufficient clinical specificity to characterize the complexity of patients’ conditions. Finally, the ICD-9-CM was outdated. Mortality statistics in the US have been organized using ICD-10 ever since 1999 and, by 2002, several other countries had already adopted the ICD-10.
The ICD-10 represents a considerable increase in the volume and diversity of diagnostic codes. In fact, it increased the number from approximately 18,000 to around 140,000. While this increase was a response to the complexity of patient conditions, it was also considered excessive by some. Critics argued that the expansion added to the difficulty of charting and billing processes, increasing costs while taking attention away from the quality of care. The controversy and debate led Congress to postpone and even consider abandoning the conversion to ICD-10.
Any doubts about the diversity delivered by so many codes? Just check out this list of 16 of the most hilarious ICD-10 codes (W220.2XD is a personal favorite; R46.1 seems a little judgy).
Finally, the US made the official transition from ICD-9 to ICD-10 on October 1, 2015. Even now, with the change long past, experts are still testing and debating over several controversies. One of these questions that couldn’t be fully answered until after implementation is the potential for improvement or change in clinical data quality caused by this new system.
Health policy research to the rescue
As previously mentioned, researchers from the Dartmouth Institute for Health Policy & Clinical Practice recently published a study in Medical Care, which attempted to quantify the continuity of disease rates during the transition from ICD-9 to -10. They accessed 100% of the inpatient Medicare data from 2012-2015. They organized diagnostic groups using standardized code sets, and the transition between 9 and 10 was mapped using three different methods. The initial test compared the annual changes from the third (Q3) to the fourth quarter (Q4) across each year.
What Mainor et al. found was all over the map. They found “substantial discontinuity” in certain conditions, while they found others more or less stable (see Figure 1 below for three such examples). Using the translation tool from Chronic Conditions Data Warehouse (CCW), they identified an immediate increase in chronic lung disease of 3.2%. The prevalence of depression dropped 1.8% from Q4 2014 to Q4 2015. In fact, the rate of depression dropped more than 2% from the week before the transition to the week after. These findings were substantially greater than the disruption initially reported by the CCW [pdf]. Yet myocardial infarction rates didn’t flinch.
Figure 1 from Mainor et al. The epitome of what the internet calls ‘gifs that end too soon.’
Variability in variations
The discontinuities are confusing to interpret. The authors point out that there was no “detectable, systematic difference” in the conditions documented with either set of codes. Even asking clinicians to look into the differences did not produce any additional clarity. Instead, the authors point to a combination of influences:
- Assumptions and trade-offs built into the software they used to translate code sets,
- Level of detail and other organizational differences, interfering with the precision of cross-mapping diagnoses, and
- Coding errors by clinicians and coders, from the novelty of ICD-10 codes.
One major limitation of the analysis is the follow-up time available after the transition period. The ICD-10 transition occurred in October 2015, and the study only extends through the end of 2015. I’d like to see what happens after the turbulence captured (Figure 1). Does it eventually level out? Do the changes regress back to previous levels or settle in at these new rates? It certainly looks as if the prevalence rate of chronic lung disease may be returning to the baseline observed before the transition.
On the plus side, the study was using the entire Medicare inpatient dataset, which means sampling bias was minimized. They used three different systems for grouping codes into diagnostic clusters and translating them between versions 9 and 10, which means the turbulence is pretty robust. Among the many Bradford Hill criteria addressed here, there is undoubtedly a clear temporal association identified.
Changing measurements changes measures
These changes are definitely concerning for several reasons. Payers, administrators, and policymakers all rely on these trends to make program decisions and allocate resources. This research raises the possibility that discontinuities in these numbers could lead to shifts in appropriations or reimbursement decisions.
Finding ways to quantify and adjust for discontinuities is critical. It’s certainly possible that these discontinuities represent a lack of familiarity on the part of coders with the new codes. If so, the coding errors could either correct or stabilize over time. On the other hand, the new sets of codes themselves may have fundamentally changed who gets counted. ICD-10 codes may be more or less generously defined for specific conditions. (Sidebar: Chris Endres’s free, searchable tabular index of ICD-9 codes has been a true companion for years. In researching for this post, I was also happy to notice the launch of a similar ICD-10-CM tabular index!)
We still don’t know whether these changes represent a correction to the biased code applications under ICD-9 or a new bias. Or both. And we don’t know whether these new incidence rates will stabilize over time. Regardless, researchers will also want to take a hard look at any plans to study longitudinal trends based on ICD codes that include the transition.
Ben King