Do You Speak DRG ?

Patient Classification Systems are constructed by collapsing similar diagnoses and/or treatments into not too many discrete groups, typically about 500 to 1000. These patient categories are usable if their labels are comprehensible.
In this study, the complexity of the DRG labels of several DRG systems is explored over time.

DRG systems used in the USA (Medicare DRG 1994 until 2008), in Australia (AN-DRG 3.1, AR-DRG 4.1, AR-DRG 5.2), and in Germany (G-DRG 2003 until 2008) were inspected. As data, only the DRG labels had to be looked at.

For each DRG label, the following conjunctions were counted: "and", "or", "with", "without", "except"; commas, slashes, pairs of brackets and the word "age". Sequenced conjunctions are counted as 1.

Whereas the number of conjunctions used in labels of Medicare DRGs stays relatively unchanged over time (the quota of labels with four or more conjunctions increases from 9 % to 11 %), there is some change within the labels from AN-DRG 3.1 through AR-DRG 4.1 to AR-DRG 5.2 (11 / 15 / 13 %). In contrast, there is a great increase in the use of conjunctions in the G-DRG labels (17 % to 49 %).

The use of conjunctions does not only impair the comprehensibility of DRG labels but sometimes makes expressions ambiguous or practically unintellegible.
To get better DRG labels, it could be of help to use logical brackets, to split up complex DRGs for clinical homogeneity, or to construct patient classification systems for use in "cushioned" remuneration systems.

A DRG system should not only be useful for statisticians, but it should also serve as a basis for discussions between clinicians and managers. Thus it is necessary to make sure that DRG labels are compehensible. This is especially important when DRG systems are strongly optimised with regard to statistical performance (as was done with the G-DRG system).

When you learn a new language, it will be easier if you already know some similar words from other languages you have learnt before. And you are more at ease if the new language does not use excessively complex structures. – A DRG system1 contains a list of labels of patient categories (DRGs) which are offered to be used during discussions between managers and clinicians. This offer will be accepted as a matter of course the more the words used within the DRG labels are comprehensible and the more the structure of the DRG labels is not excessively complex.

Patient Classification Systems are constructed by collapsing similar diagnoses and/or treatments within not too many discrete groups, typically about 500 to 1000. The aim was to get a handy and comprehensible list of names of clinical treatments that could be used as a basis for managerial discussions and decisions of clinicians and managers. Furthermore, the patient categories should serve as "production units" within a remuneration system.

Since the beginning of DRG history, it was always a challenge to get a list with not too many entries. Patient categories should be clinically meaningful. And from the economic viewpoint they should be expected to generate similar costs.

The general approach was to define a list of "Base DRGs" ("Adjacent DRGs") and additional criteria such as age and lists of complications and comorbidities ("CC lists") which could serve to split Base DRGs into subgroups which are statistically as consistent as possible.

Two approaches to construct subgroups (named "DRGs" were developed: the first was to split only selected Base DRGs into statistically different subgroups. This resulted in DRG lists of typically 500 to 700 DRGs.2

The second approach was to split all Base DRGs according to the same criteria. Normally, the selected criteria created several split levels according to the severity of the diagnoses. This resulted in systems with typically more than 1000 DRGs.3

The statistical goodness of the resulting patient categories ("DRGs") is usually expressed by measuring the explained proportion of variance of the dependent variable (usually length of stay or treatment costs) in comparison to the variance which results when the patients are not categorised. It is called "variance reduction" (R²). Its value ranges from zero (the groups do not explain any variance) to 1 (the groups are able to explain the full variance).

The first DRG systems achieved R² values ranging from 10 % to 30 %, sometimes 40 % or later even nearly 50 %. From a statistical viewpoint, these were not overwhelming results. But by using more exact methods of cost calculation and exploiting all statistical possibilities, one got higher R² values.

[Table 1] An exceptionally estonishing example is the G-DRG system: the version of 2003 was a mapping of the Australian AR-DRG system 4.1. At that time, with 664 DRGs, a R² of 46.6 % was achieved (trimmed: 62.3 %). The actual version for 2007 comprises 1082 DRGs and explains 70.7 % (trimmed: 80.5 %) of the variance. This is a value which is high as one only dreamed at before.

In this paper, I want to look into one side-effect of this statistically excellent result: what effects did this refinement work have on the comprehensibilty of the DRG labels?

The following DRG systems that are used in the USA (Medicare DRG 1993/94 until 2007), in Australia (ANDRG-3.1, ARDRG-4.1, ARDRG-5.2), and in Germany (GDRG-2003 until 2007) were inspected. As data, only the DRG labels had to be looked at. These were available on the internet.4

One aspect of comprehensibility is the logical structure of a statement. In the case of DRG labels, a question can be: how many different expressions for diagnoses and other information are put together? The hypothesis is that each block of words between conjunctions (i. e. words which join terms) can be understood as a logical (and comprehensible) unit. The formal complexity of a DRG label can then be counted by the number of the conjunctions used.

The conjunctions which were counted for each DRG label, were: "and", "or"5, "with", "without", "except", commas, slashes, pairs of brackets and the word "age".6 If conjunctions were sequenced, they were counted as one only conjunction.

A little computer program was written to count these words.7

[Table 2] The graphic above shows on the first line of graphic panels the growth of the total numbers of DRGs, of conjunctions and of characters used for the DRG labels. On the second line of panels the growth of the average numbers of conjunctions and of characters is shown.

It is remarkable that the change of these numbers for the Australian DRGs (labelled as «AR») is negligible compared to the growth of the same numbers for the German DRGs.

Medicare DRGs shows a very little growth until 2007. A visible change went with the shift to the refined MSDRG-System in 2008.

[Table 3] The next graphic shows the total number of DRGs and the use of conjunctions within American, Australian and German DRG systems over time. The total number of DRGs is displayed by the length of the bars. The bars are split into fields whose relative sizes represent the number of DRGs with a count of conjunctions of 0, 1, 2, . . . This number of conjunctions is written below the bars. The number of DRGs with the same count of conjunctions is shown above the bars. The fields for DRGs with up to tree conjunctions are coloured in blue. The fields for DRGs with four or more conjunctions are coloured in orange and red. Thus the evolution of the relation of the number of DRGs with less than four and with four or more conjunctions can be seen at first glance.

It can be seen, that the lengths of the bars (and thus the total numbers of DRGs) remain nearly unchanged for Medicare DRGs and Australian DRGs. Furthermore, the relative proportions of the light blue and the orange fields (thus the numbers of DRGs with three, four conjunctions and more) remain rather stable with Medicare DRGs but increase much with German DRGs. The number of GDRGs with zero, one or two conjunctions (the total length of the first three fields) has even decreased since 2004, although the total number of GDRGs has increased.

The first DRG system used for remuneration was Medicare DRG 2.0 (1983). It had 470 DRGs. Ten years later, the number of groups was still below 500. The 2007 version encompasses 538 groups. Thus the increase of the number of Medicare DRGs was rather modest.

[Table 4] The quota of labels with four conjunctions or more increased from 1992 until 2006 from 9 % to 10 %). Only 13 DRG labels (2.4 %) show five or more conjunctions. Thus most of the DRG labels use three or fewer conjunctions.

The shift to MS-DRG for 2008 (with the introduction of MCCs) brought a serious change: The number of DRGs without a conjunction sank from 84 DRGs to 44 DRGs, whereas the number of DRGs with two conjunctions increased from 126 to 263, and the number of DRGs with three conjunctions increased from 76 to 162. But because the total number also increased (from 538 to 745 DRGs), the quota of labels with four conjunctions or more increased only from 10 % to 13 %.

[Table 5] (Preliminary remark for Table 5: When we look at the examples of complex Medicare DRG labels, we see immediately that an additional problem arises for people who are not too familiar with medical expressions: words within long DRG labels are abbreviated until they are no longer recognizable.8 This problem will not be treated in this paper.)

The most complex DRG label was found for Medicare DRG 541 (version 24 for 2007). Eight conjunctions were used (and many abbreviations!). The full label is: "ECMO or Tracheostomy With Mechanical Ventilation 96+ Hours or Principal Diagnosis Except Face, Mouth and Neck With Major O.R. Procedure". (From a purely logical point of view it is not clear if the "or" of the part "or Principal Diagnosis Except . . ." corresponds to the "or" of "or Tracheostomy" or to the "With" of "With Mechanical Ventilation 96+ Hours". If there is some knowlegde of the matter, then the question can be decided in favour of the second possibility.)

The AN-DRG system 3.1 (with 667 DRGs) was used in 1996. The refined AR-DRG system 4.0 (with 661 DRGs) was introduced in 1998. The currently used system ARDRG-5.2 encompasses 665 DRGs.

[Table 6] The quota of labels with four conjunctions or more changed from 1996 through 1998 until 2006 from 11 % through 15 % to 13 %. 28 DRG labels (4.2 %) show five or more conjunctions. Thus many DRG labels use three or fewer conjunctions.

[Table 7] The most complex DRG labels were found for the ARDRGs belonging to Base AR-DRG I75: Seven conjunctions were used to define each of these three ARDRGs.

To be able to understand the logical structure of the labels, logical brackets are used where necessary (e. g. for ARDRG H02B, H61B and I73B, shown in the tables above).

While the G-DRG system started with 661 DRGs for the year 2003; the G-DRG system 2008 has arrived at 1137 DRGs.

[Table 8] The quota of labels with four or more conjunctions increased from 2003 until 2008 from 17 % to 49 %. 367 G-DRG labels (32 %) show five or more conjunctions. Thus nearly half of all DRG labels use four or more conjunctions, and nearly one third use five or more!

I will take B02B as a first example of a complex G-DRG label. Word-by-word translations are made by the author (with commas set according to the German original).

[Table 9] In the 2007 version this G-DRG is described with the use of nine conjunctions. (Although the resulting label is very complicated, there are G-DRG labels with even more conjunctions!) We can first see a list of three groups of procedures (complex craniotomy; spinal procedures; other expensive nervous procedures). Then three conditions are added (with motor ventilation over 95 hours; without radiotherapy with more than eight sessions; with age < 6 years). What follows is rather unclear. I interpret it as an alternative age limit (age < 18 years): if this limit is fullfilled, the next two conditions have to be coded (with major intracranial procedure; with catastrophic CC). But now, the question will arise: why is motor ventilation named a second time? Is there another logic than the one described above? Is it not correct to think in terms of a second, age-dependent branch?9 If one reads Bartkowski et al.10 one is astonished once more: he says that according to his analyses, the motor ventilation split does not have any relevance in this DRG because cases with such ventilations will be assigned to one of the Pre-MDC Base GDRGs (A06 to A13).

When compared with the G-DRG label in the sequence of the years since 2003, one sees how the texts have grown, how the conjunctions have been multiplying, how it takes more and more effort to decipher the meaning.

[Table 10] Seen in the context of the other DRGs of Base DRG B02, it becomes a challenge to find out the real differences between these DRGs in such a manner that one is able to explain them to someone else (and there will be hardly a difference if he is a clinician or a manager).

[Table 11] A second example: the Leukaemia Base G-DRG R60 is differentiated into seven GDRGs (R60A until RA60G) according to severity of illness and intensity of treatment. A look at the label list shows that there are rather easily understandable labels with three or fewer conjunctions. But others have nine or ten conjunctions. It is rather hard to understand their meaning by reading them only once.

[Table 12] To understand the labels better, they were subjected to a logical analsis. A certain system could be recognised throughout the DRGs of this Base G-DRG, yet I am still unable to assign quickly a patient which belongs to Base G-DRG R60, to the correct G-DRG.11

[Table 13] The last table shows that the above two examples are not rare cases, but that there is a rather long list of GDRGs which use many conjunctions and which are therefore hard to understand.

We have seen that the use of conjunctions impairs the comprehensibility of DRG labels. The effort required to ascertain the meaning of a logically complex DRG label can be rather great.

The method chosen was a simple one. It can be criticised that not all conjunctions used for the analysis have the same impact on comprehensibility. For instance, the use of commas is easier to understand than a sequence of and's and or's. Yet it was helpful to detect quite an interesting set of problematic DRG labels.

The conjunctions "more", "less", "<", ">" were not counted because their relation to the previous words is always strong and the following words (a number and a unit) are much easier to understand than clinical terms.

The overuse of conjunctions led to the following types of problems:

• The logic of sequences of commas, and's and or's is not anymore immediately comprehensible.12
• The logic can become ambiguous.13
• The logic can even lead to practically unintellegible labels.14
• Usability is impaired by concatenating different clinical contents which are expected to have the same costs, into a single DRG.15

Other techniques which generate not easily understandable labels are:

• The creation and use of new and special abbreviations.16
• The use of differentiating terms in only a part of all DRGs of a Base DRG.17
• The use of container words like "certain OR procedures".18
• The overuse of "hotchpotch" DRGs: DRGs for "other" diagnoses.19
• The distribution of DRGs which belong to one single Base DRG, into several Base DRGs.20

To get better DRG labels, it could be of help:

• To use logical brackets (as it was done in the AR-DRG system).
• To split up complex DRGs for clinical homogeneity.21
• To construct patient classification systems for the use within "cushioned" remuneration systems. (For instance: define for each DRG combined per diem and per case rates22 or create systems in which more than one group can be assigned to each case.23)

A DRG system should not only be useful for statisticians, but it should also serve as a basis for discussions between clinicians and managers. Thus it is necessary to make sure that DRG labels are compehensible. This is especially important when DRG systems are strongly optimised with regard to statistical performance (as it was done with the G-DRG system).